Re: [REGRESSION 4.20-rc1] 45975c7d21a1 ("rcu: Define RCU-sched API in terms of RCU for Tree RCU PREEMPT builds")

[Paul E. McKenney]

On Wed, Nov 14, 2018 at 12:20:13PM -0800, Paul E. McKenney wrote:
> On Tue, Nov 13, 2018 at 07:10:37AM -0800, Paul E. McKenney wrote:
> > On Tue, Nov 13, 2018 at 03:54:53PM +0200, Ville Syrjälä wrote:
> > > Hi Paul,
> > > 
> > > After 4.20-rc1 some of my 32bit UP machines no longer reboot/shutdown.
> > > I bisected this down to commit 45975c7d21a1 ("rcu: Define RCU-sched
> > > API in terms of RCU for Tree RCU PREEMPT builds").
> > > 
> > > I traced the hang into
> > > -> cpufreq_suspend()
> > >  -> cpufreq_stop_governor()
> > >   -> cpufreq_dbs_governor_stop()
> > >    -> gov_clear_update_util()
> > >     -> synchronize_sched()
> > >      -> synchronize_rcu()
> > > 
> > > Only PREEMPT=y is affected for obvious reasons, but that couldn't
> > > explain why the same UP kernel booted on an SMP machine worked fine.
> > > Eventually I realized that the difference between working and
> > > non-working machine was IOAPIC vs. PIC. With initcall_debug I saw
> > > that we mask everything in the PIC before cpufreq is shut down,
> > > and came up with the following fix:
> > > 
> > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> > > index 7aa3dcad2175..f88bf3c77fc0 100644
> > > --- a/drivers/cpufreq/cpufreq.c
> > > +++ b/drivers/cpufreq/cpufreq.c
> > > @@ -2605,4 +2605,4 @@ static int __init cpufreq_core_init(void)
> > >         return 0;
> > >  }
> > >  module_param(off, int, 0444);
> > > -core_initcall(cpufreq_core_init);
> > > +late_initcall(cpufreq_core_init);
> > 
> > Thank you for testing this and tracking it down!
> > 
> > I am glad that you have a fix, but I hope that we can arrive at a less
> > constraining one.
> > 
> > > Here's the resulting change in inutcall_debug:
> > >   pci 0000:00:00.1: shutdown
> > >   hub 4-0:1.0: hub_ext_port_status failed (err = -110)
> > >   agpgart-intel 0000:00:00.0: shutdown
> > > + PM: Calling cpufreq_suspend+0x0/0x100
> > >   PM: Calling mce_syscore_shutdown+0x0/0x10
> > >   PM: Calling i8259A_shutdown+0x0/0x10
> > > - PM: Calling cpufreq_suspend+0x0/0x100
> > > + reboot: Restarting system
> > > + reboot: machine restart
> > > 
> > > I didn't really look into what other ramifications the cpufreq
> > > initcall change might have. cpufreq_global_kobject worries
> > > me a bit. Maybe that one has to remain in core_initcall() and
> > > we could just move the suspend to late_initcall()? Anyways,
> > > I figured I'd leave this for someone more familiar with the
> > > code to figure out ;) 
> > 
> > Let me guess...
> > 
> > When the system suspends or shuts down, there comes a point after which
> > there is only a single CPU that is running with preemption and interrupts
> > are disabled.  At this point, RCU must change the way that it works, and
> > the commit you bisected to would make the change more necessary.  But if
> > I am guessing correctly, we have just been getting lucky in the past.
> > 
> > It looks like RCU needs to create a struct syscore_ops with a shutdown
> > function and pass this to register_syscore_ops().  Maybe a suspend
> > function as well.  And RCU needs to invoke register_syscore_ops() at
> > a time that causes RCU's shutdown function to be invoked in the right
> > order with respect to the other work in flight.  The hope would be that
> > RCU's suspend function gets called just as the system transitions into
> > a mode where the scheduler is no longer active, give or take.
> > 
> > Does this make sense, or am I confused?
> 
> Well, it certainly does not make sense in that blocking is still legal
> at .shutdown() invocation time, which means that RCU cannot revert to
> its boot-time approach at that point.  Looks like I need hooks in a
> bunch of arch-dependent functions.  Which is certainly doable, but will
> take a bit more digging.

A bit more detail, after some additional discussion at Linux Plumbers
conference...

The preferred approach is to hook into syscore_suspend(),
syscore_resume(), and syscore_shutdown().  This can be done easily by
creating an appropriately initialized struct syscore_ops and passing a
pointer to it to register_syscore_ops() during boot.  Taking these three
functions in turn:

syscore_suspend():

o	arch/x86/kernel/apm_32.c suspend(), standby()

	These calls to syscore_suspend() has interrupts disabled, which
	is very good, but they are immediately re-enabled, and only then
	is the call to set_system_power_state().  Unless both interrupts
	and preemption are prevented somehow, it is not safe for
	CONFIG_PREEMPT=y RCU implementations to revert back to boot-time
	behavior at this point.

o	drivers/xen/manage.c xen_suspend()

	This looks to have interrupts disabled throughout.  It is also
	invoked within stop_machine(), which means that the other CPUs,
	though online, are quiescent.  This allows RCU to safely switch
	back to early boot operating mode.  That is, this is safe only
	if there is no interaction with RCU-preempt read-side critical
	sections that might well be underway in the other CPUs.  This
	assumption is likely violated in CONFIG_PREEMPT=y kernels.  One
	alternative that would work with RCU in CONFIG_PREEMPT=y kernels
	is CPU-hotplug removing all but one CPU, but that might have
	some other disadvantages.

o	kernel/kexec_core.c kernel_kexec()

	Before we get here, disable_nonboot_cpus() has been invoked, which
	in turn invokes freeze_secondary_cpus(), which offlines all but
	the boot CPU.  Prior to that, all user-space tasks are frozen.
	So in this case, it would be safe for RCU to revert back to its
	boot-time behavior.  Aside from the possibility of unfreezable
	kthreads being preempted within RCU-preempt read-side critical
	sections, anyway...  :-/

	However, one can argue that as long as the kthreads preempted
	within an RCU-preempt read-side critical section are guaranteed
	to never ever run again, we might be OK.  And this guarantee
	seems consistent with the kernel_kexec() operation.  At least
	when there are no errors that cause the kernel_kexec() to return
	control to the initial kernel image...

	Of course, this line of reasoning does not apply when the
	kernel is to resume on the same hardware, as in some of the
	cases above.

o	kernel/power/hibernate.c create_image()

	Same as for kernel_kexec(), except that freeze_kernel_threads()
	is invoked, which hopefully gets all tasks out of RCU read-side
	critical sections.  So this one might actually permit RCU to
	revert back to boot-time behavior.  Except for the possibility of
	an error condition forcing an abort back into the original kernel
	image, which again could have trouble with kthreads that were
	preempted within an RCU read-side critical section throughout.

o	kernel/power/hibernate.c resume_target_kernel()
	kernel/power/hibernate.c hibernation_platform_enter()
	kernel/power/suspend.c suspend_enter()

	Same as for kernel_kexec(), but no obvious pretense of freezing
	any tasks.


syscore_resume():

o	arch/x86/kernel/apm_32.c suspend(), standby()

	Resume-time counterparts to the calls to syscore_suspend() called
	out above, with the same interrupt-enabling problem, as well as
	issues with tasks being preempted throughout within RCU-preempt
	read-side critical sections.

o	drivers/xen/manage.c xen_suspend()

	Resume-time counterpart to the calls to xen_suspend() called out
	above, with the same issues with tasks being preempted throughout
	within RCU-preempt read-side critical sections.

o	kernel/kexec_core.c kernel_kexec()

	Resume-time counterpart to the calls to kernel_kexec() called out
	above.  This is the error case that causes trouble due to the
	possibility of preempted RCU read-side critical sections.

o	kernel/power/hibernate.c create_image()
	kernel/power/hibernate.c resume_target_kernel()
	kernel/power/hibernate.c hibernation_platform_enter()
	kernel/power/hibernate.c suspend_enter()

	Resume-time counterparts to calls within kernel/power/hibernate.c
	and kernel/power/suspend.c called out above.  This is the error
	case that causes trouble due to the possibility of preempted
	RCU read-side critical sections.


syscore_shutdown():

o	kernel/reboot.c kernel_restart()
	kernel/reboot.c kernel_halt()
	kernel/reboot.c kernel_power_off()

	These appears to leave all CPUs online, which prevents RCU from
	safely reverting back to boot-time mode.


So what is to be done?

Here are the options I can see:

1.	Status quo, which means that synchronize_rcu() and friends
	cannot be used in syscore_suspend(), syscore_resume(), and
	syscore_shutdown() callbacks.  At the moment, this appears to
	be the only workable approach, though ideas and suggestions are
	quite welcome.

2.	Make each code path to syscore_suspend(), syscore_resume(), and
	syscore_shutdown() offline all but the boot CPU, ensure that
	all tasks exit any RCU read-side critical sections that they
	might be in, then run the remainder of the code path on the
	boot CPU with interrupts disabled.

	Making all tasks exit any RCU read-side critical sections is
	easy when CONFIG_PREEMPT=n via things like stop-machine, but
	it is difficult and potentially time-consuming for
	CONFIG_PREEMPT=y kernels.

3.	Do error checking so that there cannot possibly be failures
	beyond the time that syscore_suspend(), syscore_resume(),
	and syscore_shutdown() are invoked.  This is fine for
	syscore_shutdown() and syscore_resume(), but syscore_suspend()'s
	callbacks are permitted to return errors that force suspend
	failures.

	And there are syscore_suspend() callbacks that actually do
	return errors, for example, fsl_lbc_syscore_suspend()
	in arch/powerpc/sysdev/fsl_lbc.c can return -ENOMEM.
	As can save_ioapic_entries() in arch/x86/kernel/apic/io_apic.c
	and arch/x86/include/asm/io_apic.h.  And mvebu_mbus_suspend()
	in drivers/bus/mvebu-mbus.c.  And iommu_suspend() in
	drivers/iommu/intel-iommu.c.

	And its_save_disable() in drivers/irqchip/irq-gic-v3-its.c
	can return -EBUSY.

	Perhaps these can be remedied somehow, but unless that can
	be done, this approach cannot work.

4.	Your idea here!!!

							Thanx, Paul