Linux Virtualization - Apr 2016 - [PATCH v3 5/6] virt, sched: add cpu pinning to smp_call_sync_on_phys_cpu()

On Fri, Apr 01, 2016 at 09:14:33AM +0200, Juergen Gross
wrote:
> --- a/kernel/smp.c
> +++ b/kernel/smp.c
> @@ -14,6 +14,7 @@
>  #include <linux/smp.h>
>  #include <linux/cpu.h>
>  #include <linux/sched.h>
> +#include <linux/hypervisor.h>
>  
>  #include "smpboot.h"
>  
> @@ -758,9 +759,14 @@ struct smp_sync_call_struct {
>  static void smp_call_sync_callback(struct work_struct *work)
>  {
>  	struct smp_sync_call_struct *sscs;
> +	unsigned int cpu;
>  
>  	sscs = container_of(work, struct smp_sync_call_struct, work);
> +	cpu = get_cpu();
> +	hypervisor_pin_vcpu(cpu);
>  	sscs->ret = sscs->func(sscs->data);
> +	hypervisor_pin_vcpu(-1);
> +	put_cpu();
>  
>  	complete(&sscs->done);
>  }
So I don't really like this; it adds the requirement that the function
cannot schedule, which greatly limits the utility of the construct. At
this point you might as well use the regular IPI stuff.

You can easily avoid this constraint by using:

	hypervisor_pin_vcpu(smp_processor_id());

Also, for the vpinning stuff, the UP version below is sufficient, even
on SMP systems (with the current !preempt constraint). Which seems to
suggest we're not having the right interface for this.

So I would propose you add:

	smp_call_on_cpu()

As per patch 2. No promises about physical or anything. This means it
can be used freely by anyone that wants to run a function on another
cpu -- a much more useful thing.

And then build a phys variant on top.

> diff --git a/kernel/up.c b/kernel/up.c
> index afd395c..725ec44 100644
> --- a/kernel/up.c
> +++ b/kernel/up.c
> @@ -6,6 +6,7 @@
>  #include <linux/kernel.h>
>  #include <linux/export.h>
>  #include <linux/smp.h>
> +#include <linux/hypervisor.h>
>  
>  int smp_call_function_single(int cpu, void (*func) (void *info), void
*info,
>  				int wait)
> @@ -85,9 +86,17 @@ EXPORT_SYMBOL(on_each_cpu_cond);
>  
>  int smp_call_sync_on_phys_cpu(unsigned int cpu, int (*func)(void *), void
*par)
>  {
> +	int ret;
> +
>  	if (cpu != 0)
>  		return -EINVAL;
>  
> -	return func(par);
> +	preempt_disable();
> +	hypervisor_pin_vcpu(0);
> +	ret = func(par);
> +	hypervisor_pin_vcpu(-1);
> +	preempt_enable();
> +
> +	return ret;
>  }
>  EXPORT_SYMBOL_GPL(smp_call_sync_on_phys_cpu);
> -- 
> 2.6.2
>

On 01/04/16 09:43, Peter Zijlstra wrote:
> On Fri, Apr 01, 2016 at 09:14:33AM +0200, Juergen Gross wrote:
>> --- a/kernel/smp.c
>> +++ b/kernel/smp.c
>> @@ -14,6 +14,7 @@
>>  #include <linux/smp.h>
>>  #include <linux/cpu.h>
>>  #include <linux/sched.h>
>> +#include <linux/hypervisor.h>
>>  
>>  #include "smpboot.h"
>>  
>> @@ -758,9 +759,14 @@ struct smp_sync_call_struct {
>>  static void smp_call_sync_callback(struct work_struct *work)
>>  {
>>  	struct smp_sync_call_struct *sscs;
>> +	unsigned int cpu;
>>  
>>  	sscs = container_of(work, struct smp_sync_call_struct, work);
>> +	cpu = get_cpu();
>> +	hypervisor_pin_vcpu(cpu);
>>  	sscs->ret = sscs->func(sscs->data);
>> +	hypervisor_pin_vcpu(-1);
>> +	put_cpu();
>>  
>>  	complete(&sscs->done);
>>  }
> 
> So I don't really like this; it adds the requirement that the function
> cannot schedule, which greatly limits the utility of the construct. At
> this point you might as well use the regular IPI stuff.
Main reason for disabling preemption was to avoid any suspend/resume
cycles while vcpu pinning is active.

With the switch to workqueues this might not be necessary, if I've read
try_to_freeze_tasks() correctly. Can you confirm, please?
> You can easily avoid this constraint by using:
> 
> 	hypervisor_pin_vcpu(smp_processor_id());
> 
> Also, for the vpinning stuff, the UP version below is sufficient, even
> on SMP systems (with the current !preempt constraint). Which seems to
> suggest we're not having the right interface for this.
> 
> So I would propose you add:
> 
> 	smp_call_on_cpu()
> 
> As per patch 2. No promises about physical or anything. This means it
> can be used freely by anyone that wants to run a function on another
> cpu -- a much more useful thing.
Okay.
> And then build a phys variant on top.
Hmm, I'm not sure I understand what you are suggesting here.

Should this phys variant make use of smp_call_on_cpu() via an
intermediate function called on the dedicated cpu which is doing the
pinning and calling the user function then?

Or do you want the phys variant to either use smp_call_on_cpu() or to
do the pinning and call the user function by itself depending on the
environment (pinning supported)?


Juergen

On Fri, Apr 01, 2016 at 10:28:46AM +0200, Juergen Gross
wrote:
> On 01/04/16 09:43, Peter Zijlstra wrote:
> > On Fri, Apr 01, 2016 at 09:14:33AM +0200, Juergen Gross wrote:
> >> --- a/kernel/smp.c
> >> +++ b/kernel/smp.c
> >> @@ -14,6 +14,7 @@
> >>  #include <linux/smp.h>
> >>  #include <linux/cpu.h>
> >>  #include <linux/sched.h>
> >> +#include <linux/hypervisor.h>
> >>  
> >>  #include "smpboot.h"
> >>  
> >> @@ -758,9 +759,14 @@ struct smp_sync_call_struct {
> >>  static void smp_call_sync_callback(struct work_struct *work)
> >>  {
> >>  	struct smp_sync_call_struct *sscs;
> >> +	unsigned int cpu;
> >>  
> >>  	sscs = container_of(work, struct smp_sync_call_struct, work);
> >> +	cpu = get_cpu();
> >> +	hypervisor_pin_vcpu(cpu);
> >>  	sscs->ret = sscs->func(sscs->data);
> >> +	hypervisor_pin_vcpu(-1);
> >> +	put_cpu();
> >>  
> >>  	complete(&sscs->done);
> >>  }
> > 
> > So I don't really like this; it adds the requirement that the
function
> > cannot schedule, which greatly limits the utility of the construct. At
> > this point you might as well use the regular IPI stuff.
> 
> Main reason for disabling preemption was to avoid any suspend/resume
> cycles while vcpu pinning is active.
> 
> With the switch to workqueues this might not be necessary, if I've read
> try_to_freeze_tasks() correctly. Can you confirm, please?
This is not something we should worry about; the caller should ensure
the CPU stays valid; typically I would expect a caller to do
get_online_cpus() before 'computing' what CPU to send the function to.
> > So I would propose you add:
> > 
> > 	smp_call_on_cpu()
> > 
> > As per patch 2. No promises about physical or anything. This means it
> > can be used freely by anyone that wants to run a function on another
> > cpu -- a much more useful thing.
> 
> Okay.
> 
> > And then build a phys variant on top.
> 
> Hmm, I'm not sure I understand what you are suggesting here.
> 
> Should this phys variant make use of smp_call_on_cpu() via an
> intermediate function called on the dedicated cpu which is doing the
> pinning and calling the user function then?
> 
> Or do you want the phys variant to either use smp_call_on_cpu() or to
> do the pinning and call the user function by itself depending on the
> environment (pinning supported)?
Yeah, uhmm.. not sure on the details; my brain is having a hard time
engaging this morning.

Maybe just make the vpin thing an option like:

	smp_call_on_cpu(int (*func)(void *), int phys_cpu);

Also; is something like the vpin thing possible on KVM? because if we're
going to expose it to generic code like this we had maybe look at wider
support.