[PATCH v02] powerpc/pseries: Check for ceded CPU's during LPAR migration

[Michael Bringmann]

See below.

On 1/31/19 3:53 PM, Michael Bringmann wrote:
> On 1/30/19 11:38 PM, Michael Ellerman wrote:
>> Michael Bringmann <mwb at linux.vnet.ibm.com> writes:
>>> This patch is to check for cede'ed CPUs during LPM.  Some extreme
>>> tests encountered a problem ehere Linux has put some threads to
>>> sleep (possibly to save energy or something), LPM was attempted,
>>> and the Linux kernel didn't awaken the sleeping threads, but issued
>>> the H_JOIN for the active threads.  Since the sleeping threads
>>> are not awake, they can not issue the expected H_JOIN, and the
>>> partition would never suspend.  This patch wakes the sleeping
>>> threads back up.
>>
>> I'm don't think this is the right solution.
>>
>> Just after your for loop we do an on_each_cpu() call, which sends an IPI
>> to every CPU, and that should wake all CPUs up from CEDE.
>>
>> If that's not happening then there is a bug somewhere, and we need to
>> work out where.

>From Pete Heyrman:
    Both sending IPI or H_PROD will awaken a logical processors that has ceded.
    When you have logical proc doing cede and one logical proc doing prod or IPI
    you have a race condition that the prod/IPI can proceed the cede request.
    If you use prod, the hypervisor takes care of the synchronization by ignoring
    a cede request if it was preceeded by a prod.  With IPI the interrupt is
    delivered which could then be followed by a cede so OS would need to provide
    synchronization.

Shouldn't this answer your concerns about race conditions and the suitability
of using H_PROD?

Michael

> 
> Let me explain the scenario of the LPM case that Pete Heyrman found, and
> that Nathan F. was working upon, previously.
> 
> In the scenario, the partition has 5 dedicated processors each with 8 threads
> running.
> 
>>From the PHYP data we can see that on VP 0, threads 3, 4, 5, 6 and 7 issued
> a H_CEDE requesting to save energy by putting the requesting thread into
> sleep mode.  In this state, the thread will only be awakened by H_PROD from
> another running thread or from an external user action (power off, reboot
> and such).  Timers and external interrupts are disabled in this mode.
> 
> About 3 seconds later, as part of the LPM operation, the other 35 threads
> have all issued a H_JOIN request.  Join is part of the LPM process where
> the threads suspend themselves as part of the LPM operation so the partition
> can be migrated to the target server.
> 
> So, the current state is the the OS has suspended the execution of all the
> threads in the partition without successfully suspending all threads as part
> of LPM.
> 
> Net, OS has an issue where they suspended every processor thread so nothing
> can run.
> 
> This appears to be slightly different than the previous LPM stalls we have
> seen where the migration stalls because of cpus being taken offline and not
> making the H_JOIN call.
> 
> In this scenario we appear to have CPUs that have done an H_CEDE prior to
> the LPM. For these CPUs we would need to do a H_PROD to wake them back up
> so they can do a H_JOIN and allow the LPM to continue.
> 
> The problem is that Linux has some threads that they put to sleep (probably
> to save energy or something), LPM was attempted, Linux didn't awaken the
> sleeping threads but issued the H_JOIN for the active threads.  Since the
> sleeping threads don't issue the H_JOIN the partition will never suspend.
> 
> I am checking again with Pete regarding your concerns.
> 
> Thanks.
> 
>>
>>
>>> diff --git a/arch/powerpc/include/asm/plpar_wrappers.h b/arch/powerpc/include/asm/plpar_wrappers.h
>>> index cff5a41..8292eff 100644
>>> --- a/arch/powerpc/include/asm/plpar_wrappers.h
>>> +++ b/arch/powerpc/include/asm/plpar_wrappers.h
>>> @@ -26,10 +26,8 @@ static inline void set_cede_latency_hint(u8 latency_hint)
>>>  	get_lppaca()->cede_latency_hint = latency_hint;
>>>  }
>>>  
>>> -static inline long cede_processor(void)
>>> -{
>>> -	return plpar_hcall_norets(H_CEDE);
>>> -}
>>> +int cpu_is_ceded(int cpu);
>>> +long cede_processor(void);
>>>  
>>>  static inline long extended_cede_processor(unsigned long latency_hint)
>>>  {
>>> diff --git a/arch/powerpc/kernel/rtas.c b/arch/powerpc/kernel/rtas.c
>>> index de35bd8f..fea3d21 100644
>>> --- a/arch/powerpc/kernel/rtas.c
>>> +++ b/arch/powerpc/kernel/rtas.c
>>> @@ -44,6 +44,7 @@
>>>  #include <asm/time.h>
>>>  #include <asm/mmu.h>
>>>  #include <asm/topology.h>
>>> +#include <asm/plpar_wrappers.h>
>>>  
>>>  /* This is here deliberately so it's only used in this file */
>>>  void enter_rtas(unsigned long);
>>> @@ -942,7 +943,7 @@ int rtas_ibm_suspend_me(u64 handle)
>>>  	struct rtas_suspend_me_data data;
>>>  	DECLARE_COMPLETION_ONSTACK(done);
>>>  	cpumask_var_t offline_mask;
>>> -	int cpuret;
>>> +	int cpuret, cpu;
>>>  
>>>  	if (!rtas_service_present("ibm,suspend-me"))
>>>  		return -ENOSYS;
>>> @@ -991,6 +992,11 @@ int rtas_ibm_suspend_me(u64 handle)
>>>  		goto out_hotplug_enable;
>>>  	}
>>>  
>>> +	for_each_present_cpu(cpu) {
>>> +		if (cpu_is_ceded(cpu))
>>> +			plpar_hcall_norets(H_PROD, get_hard_smp_processor_id(cpu));
>>> +	}
>>
>> There's a race condition here, there's nothing to prevent the CPUs you
>> just PROD'ed from going back into CEDE before you do the on_each_cpu()
>> call below> 
>>>  	/* Call function on all CPUs.  One of us will make the
>>>  	 * rtas call
>>>  	 */
>>> diff --git a/arch/powerpc/platforms/pseries/setup.c b/arch/powerpc/platforms/pseries/setup.c
>>> index 41f62ca2..48ae6d4 100644
>>> --- a/arch/powerpc/platforms/pseries/setup.c
>>> +++ b/arch/powerpc/platforms/pseries/setup.c
>>> @@ -331,6 +331,24 @@ static int alloc_dispatch_log_kmem_cache(void)
>>>  }
>>>  machine_early_initcall(pseries, alloc_dispatch_log_kmem_cache);
>>>  
>>> +static DEFINE_PER_CPU(int, cpu_ceded);
>>> +
>>> +int cpu_is_ceded(int cpu)
>>> +{
>>> +	return per_cpu(cpu_ceded, cpu);
>>> +}
>>> +
>>> +long cede_processor(void)
>>> +{
>>> +	long rc;
>>> +
>>> +	per_cpu(cpu_ceded, raw_smp_processor_id()) = 1;
>>
>> And there's also a race condition here. From the other CPU's perspective
>> the store to cpu_ceded is not necessarily ordered vs the hcall below.
>> Which means the other CPU can see cpu_ceded = 0, and therefore not prod
>> us, but this CPU has already called H_CEDE.
>>
>>> +	rc = plpar_hcall_norets(H_CEDE);
>>> +	per_cpu(cpu_ceded, raw_smp_processor_id()) = 0;
>>> +
>>> +	return rc;
>>> +}
>>> +
>>>  static void pseries_lpar_idle(void)
>>>  {
>>>  	/*
>>
>> cheers
>>
>>
> 

-- 
Michael W. Bringmann
Linux Technology Center
IBM Corporation
Tie-Line  363-5196
External: (512) 286-5196
Cell:       (512) 466-0650
mwb at linux.vnet.ibm.com