Re: [PATCH v12 3/3] tracing: Centralize preemptirq tracepoints and unify their usage

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Aug 08, 2018 at 11:23:09AM -0400, Steven Rostedt wrote:
> On Wed, 8 Aug 2018 08:05:58 -0700
> "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> 
> > On Wed, Aug 08, 2018 at 10:49:10AM -0400, Steven Rostedt wrote:
> > > On Wed, 8 Aug 2018 07:33:10 -0700
> > > "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> > >   
> > > > On Wed, Aug 08, 2018 at 09:07:24AM -0400, Steven Rostedt wrote:  
> > > > > On Wed, 8 Aug 2018 06:03:02 -0700
> > > > > "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> > > > >     
> > > > > >  What's wrong with a this_cpu_inc()? It's atomic for the CPU. Although    
> > > > > > > it wont be atomic for the capture of the idx. But I also don't see
> > > > > > > interrupts being disabled, thus an NMI is no different than any
> > > > > > > interrupt doing the same thing, right?      
> > > > > > 
> > > > > > On architectures without increment-memory instructions, if you take an NMI
> > > > > > between the load from sp->sda->srcu_lock_count and the later store, you
> > > > > > lose a count.  Note that both __srcu_read_lock() and __srcu_read_unlock()
> > > > > > do increments of different locations, so you cannot rely on the usual
> > > > > > "NMI fixes up before exit" semantics you get when incrementing and
> > > > > > decrementing the same location.    
> > > > > 
> > > > > And how is this handled in the interrupt case? Interrupts are not
> > > > > disabled here.    
> > > > 
> > > > Actually, on most architectures interrupts are in fact disabled:
> > > > 
> > > > #define this_cpu_generic_to_op(pcp, val, op)				\
> > > > do {									\
> > > > 	unsigned long __flags;						\
> > > > 	raw_local_irq_save(__flags);					\
> > > > 	raw_cpu_generic_to_op(pcp, val, op);				\
> > > > 	raw_local_irq_restore(__flags);					\
> > > > } while (0)
> > > > 
> > > > NMIs, not so much.  
> > > 
> > > And do these archs have NMIs?  
> > 
> > It would appear so:
> 
> Well the next question is, which of these archs that use it are in this
> list.
> 
> > $ find . -name 'Kconfig*' -exec grep -l 'select HAVE_NMI\>' {} \;
> > ./arch/sparc/Kconfig
> > ./arch/s390/Kconfig
> > ./arch/arm/Kconfig
> > ./arch/arm64/Kconfig
> > ./arch/mips/Kconfig
> > ./arch/sh/Kconfig
> > ./arch/powerpc/Kconfig
> 
> Note, I know that powerpc "imitates" an NMI. It just sets the NMI as a
> priority higher than other interrupts.

Plus as you say below, its local_inc() is atomic, and thus NMI-safe,
and thus the _nmi() approach would work.

> > ./arch/x86/Kconfig
> 
> And we get this:
> 
> $ git grep this_cpu_add_4
> arch/arm64/include/asm/percpu.h:#define this_cpu_add_4(pcp, val) _percpu_add(pcp, val)
> arch/s390/include/asm/percpu.h:#define this_cpu_add_4(pcp, val) arch_this_cpu_to_op_simple(pcp, val, +)
> arch/s390/include/asm/percpu.h:#define this_cpu_add_4(pcp, val) arch_this_cpu_add(pcp, val, "laa", "asi", int)
> arch/x86/include/asm/percpu.h:#define this_cpu_add_4(pcp, val)  percpu_add_op((pcp), val)
> 
> Which leaves us with sparc, arm, mips, sh and powerpc.
> 
> sh is almost dead, and powerpc can be fixed, which I guess leaves us
> with sparc, arm and mips.

If we want to stick with the current srcu_read_lock() and srcu_read_unlock(),
you mean?  I would like that sort of outcome, at least assuming we are not
hammering any of the architectures.

							Thanx, Paul