[next][bug] rcu_dyntick_kick_cpu() kills ARM SMP.

[Paul E. McKenney]

On Wed, Jun 29, 2011 at 09:58:50AM +0100, Marc Zyngier wrote:
> On 28/06/11 18:25, Paul E. McKenney wrote:
> > On Tue, Jun 28, 2011 at 04:08:37PM +0100, Marc Zyngier wrote:
> >> Paul,
> >>
> >> I've updated my -next tree (to next-20110628) today, and discovered 
> >> that my favorite ARM board wouldn't boot anymore:
> >>
> >> [...]
> >> Hierarchical RCU implementation.
> >> NR_IRQS:128 nr_irqs:128 128
> >> Console: colour dummy device 80x30
> >> Calibrating delay loop... 83.35 BogoMIPS (lpj=416768)
> >> pid_max: default: 32768 minimum: 301
> >> Mount-cache hash table entries: 512
> >> CPU: Testing write buffer coherency: ok
> >> Calibrating local timer... 104.04MHz.
> >> CPU1: Booted secondary processor
> >> CPU1: Unknown IPI message 0x1
> >> CPU2: Booted secondary processor
> >> CPU2: Unknown IPI message 0x1
> >> CPU3: Booted secondary processor
> >> CPU3: Unknown IPI message 0x1
> >> Brought up 4 CPUs
> >> SMP: Total of 4 processors activated (333.92 BogoMIPS).
> >> ------------[ cut here ]------------
> >> WARNING: at kernel/smp.c:320 smp_call_function_single+0xe4/0x1c0()
> >> NET: Registered protocol family 16
> >> Modules linked in:
> >> [<c00415d4>] (unwind_backtrace+0x0/0xf4) from [<c0056184>] (warn_slowpath_common+0x4c/0x64)
> >> [<c0056184>] (warn_slowpath_common+0x4c/0x64) from [<c00561b8>] (warn_slowpath_null+0x1c/0x24)
> >> [<c00561b8>] (warn_slowpath_null+0x1c/0x24) from [<c0088218>] (smp_call_function_single+0xe4/0x1c0)
> >> [<c0088218>] (smp_call_function_single+0xe4/0x1c0) from [<c0094804>] (rcu_start_gp+0x184/0x310)
> >> [<c0094804>] (rcu_start_gp+0x184/0x310) from [<c00955b0>] (__rcu_process_callbacks+0x274/0x398)
> >> [<c00955b0>] (__rcu_process_callbacks+0x274/0x398) from [<c0095708>] (rcu_process_callbacks+0x34/0x5c)
> >> [<c0095708>] (rcu_process_callbacks+0x34/0x5c) from [<c005c964>] (__do_softirq+0xa4/0x16c)
> >> [<c005c964>] (__do_softirq+0xa4/0x16c) from [<c005cc0c>] (irq_exit+0x80/0x9c)
> >> [<c005cc0c>] (irq_exit+0x80/0x9c) from [<c00353cc>] (do_local_timer+0x54/0x70)
> >> [<c00353cc>] (do_local_timer+0x54/0x70) from [<c003b618>] (__irq_svc+0x38/0xc0)
> >> Exception stack(0xdf467f90 to 0xdf467fd8)
> >> 7f80:                                     df466000 00000000 df467fd8 00000000
> >> 7fa0: df466000 c045dd24 c034f6cc 00000000 c0445514 410fb020 70409ddc 00000000
> >> 7fc0: 00000000 df467fd8 c003c4ac c003c4b0 60000013 ffffffff
> >> [<c003b618>] (__irq_svc+0x38/0xc0) from [<c003c4b0>] (default_idle+0x24/0x28)
> >> [<c003c4b0>] (default_idle+0x24/0x28) from [<c003ccd0>] (cpu_idle+0x9c/0xdc)
> >> [<c003ccd0>] (cpu_idle+0x9c/0xdc) from [<70348734>] (0x70348734)
> >> ---[ end trace 1b75b31a2719ed1c ]---
> >>
> >> ... and here it dies.
> >>
> >> The offending commit is b983032b7 (rcu: Avoid grace-period overflow for 
> >> long dyntick-idle periods). rcu_dyntick_kick_cpu() tries to do a CPU 
> >> cross-call with interrupts disabled, which kills the box. Reverting this
> >> patch results in a working system.
> > 
> > That does sound problematic...
> > 
> >> My RCU-foo being rather low, I haven't dug deeper into this. Please let
> >> me know if you want me to test anything.
> > 
> > I will put together a patch to defer the actual cross-call until irqs
> > are enabled.  The call would be from softirq -- that is OK, correct?
> 
> That should indeed fix the problem, as interrupts are normally enabled
> in softirq.

Hello, Marc,

I played with this some, but eventually chose to defer this commit to
3.1.  Thank you for testing -- and it will be back!

							Thanx, Paul