Re: [pchecks v1 2/4] Use raw cpu ops for calls that would trigger with checks

Re: [pchecks v1 2/4] Use raw cpu ops for calls that would trigger with checks

[Ingo Molnar]

(netdev Cc:-ed)

* Christoph Lameter <cl@linux.com> wrote:

> These location triggered during testing with KVM.
> 
> These are fetches without preemption off where we judged that
> to be more performance efficient or where other means of
> providing synchronization (BH handling) are available.

> Index: linux/include/net/snmp.h
> ===================================================================
> --- linux.orig/include/net/snmp.h	2013-09-12 13:26:29.216103951 -0500
> +++ linux/include/net/snmp.h	2013-09-12 13:26:29.208104037 -0500
> @@ -126,7 +126,7 @@ struct linux_xfrm_mib {
>  	extern __typeof__(type) __percpu *name[SNMP_ARRAY_SZ]
>  
>  #define SNMP_INC_STATS_BH(mib, field)	\
> -			__this_cpu_inc(mib[0]->mibs[field])
> +			raw_cpu_inc(mib[0]->mibs[field])
>  
>  #define SNMP_INC_STATS_USER(mib, field)	\
>  			this_cpu_inc(mib[0]->mibs[field])
> @@ -141,7 +141,7 @@ struct linux_xfrm_mib {
>  			this_cpu_dec(mib[0]->mibs[field])
>  
>  #define SNMP_ADD_STATS_BH(mib, field, addend)	\
> -			__this_cpu_add(mib[0]->mibs[field], addend)
> +			raw_cpu_add(mib[0]->mibs[field], addend)

Are the networking folks fine with allowing unafe operations of SNMP stats 
in preemptible sections, or should the kernel produce an optional warning 
message if CONFIG_PREEMPT_DEBUG=y and these ops are used in preemptible 
(non-bh, non-irq-handler, non-irqs-off, etc.) sections?

RAW_SNMP_*_STATS() ops could be used to annotate those places where that 
kind of usage is safe.

Thanks,

	Ingo

Re: [pchecks v1 2/4] Use raw cpu ops for calls that would trigger with checks

[Eric Dumazet]

On Tue, 2013-09-24 at 09:32 +0200, Ingo Molnar wrote:
> (netdev Cc:-ed)
> 
> * Christoph Lameter <cl@linux.com> wrote:
> 
> > These location triggered during testing with KVM.
> > 
> > These are fetches without preemption off where we judged that
> > to be more performance efficient or where other means of
> > providing synchronization (BH handling) are available.
> 
> > Index: linux/include/net/snmp.h
> > ===================================================================
> > --- linux.orig/include/net/snmp.h	2013-09-12 13:26:29.216103951 -0500
> > +++ linux/include/net/snmp.h	2013-09-12 13:26:29.208104037 -0500
> > @@ -126,7 +126,7 @@ struct linux_xfrm_mib {
> >  	extern __typeof__(type) __percpu *name[SNMP_ARRAY_SZ]
> >  
> >  #define SNMP_INC_STATS_BH(mib, field)	\
> > -			__this_cpu_inc(mib[0]->mibs[field])
> > +			raw_cpu_inc(mib[0]->mibs[field])
> >  
> >  #define SNMP_INC_STATS_USER(mib, field)	\
> >  			this_cpu_inc(mib[0]->mibs[field])
> > @@ -141,7 +141,7 @@ struct linux_xfrm_mib {
> >  			this_cpu_dec(mib[0]->mibs[field])
> >  
> >  #define SNMP_ADD_STATS_BH(mib, field, addend)	\
> > -			__this_cpu_add(mib[0]->mibs[field], addend)
> > +			raw_cpu_add(mib[0]->mibs[field], addend)
> 
> Are the networking folks fine with allowing unafe operations of SNMP stats 
> in preemptible sections, or should the kernel produce an optional warning 
> message if CONFIG_PREEMPT_DEBUG=y and these ops are used in preemptible 
> (non-bh, non-irq-handler, non-irqs-off, etc.) sections?
> 
> RAW_SNMP_*_STATS() ops could be used to annotate those places where that 
> kind of usage is safe.


I would rather not use RAW_ prefix in the macro, but add debugging
check to make sure we use _BH() variant in the right context.

BUG_ON(!in_softirq())