Re: [PATCH RFC v2 rcu 3/8] srcu: Check for consistent per-CPU per-srcu_struct NMI safety

[Frederic Weisbecker]

On Mon, Oct 03, 2022 at 04:57:18AM -0700, Paul E. McKenney wrote:
> On Mon, Oct 03, 2022 at 12:13:31PM +0200, Frederic Weisbecker wrote:
> > On Sun, Oct 02, 2022 at 04:51:03PM -0700, Paul E. McKenney wrote:
> > > On Mon, Oct 03, 2022 at 12:06:19AM +0200, Frederic Weisbecker wrote:
> > > > On Thu, Sep 29, 2022 at 11:07:26AM -0700, Paul E. McKenney wrote:
> > > > > This commit adds runtime checks to verify that a given srcu_struct uses
> > > > > consistent NMI-safe (or not) read-side primitives on a per-CPU basis.
> > > > > 
> > > > > Link: https://lore.kernel.org/all/20220910221947.171557773@xxxxxxxxxxxxx/
> > > > > 
> > > > > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > > > > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > > > > Cc: John Ogness <john.ogness@xxxxxxxxxxxxx>
> > > > > Cc: Petr Mladek <pmladek@xxxxxxxx>
> > > > > ---
> > > > >  include/linux/srcu.h     |  4 ++--
> > > > >  include/linux/srcutiny.h |  4 ++--
> > > > >  include/linux/srcutree.h |  9 +++++++--
> > > > >  kernel/rcu/srcutree.c    | 38 ++++++++++++++++++++++++++++++++------
> > > > >  4 files changed, 43 insertions(+), 12 deletions(-)
> > > > > 
> > > > > diff --git a/include/linux/srcu.h b/include/linux/srcu.h
> > > > > index 2cc8321c0c86..565f60d57484 100644
> > > > > --- a/include/linux/srcu.h
> > > > > +++ b/include/linux/srcu.h
> > > > > @@ -180,7 +180,7 @@ static inline int srcu_read_lock_nmisafe(struct srcu_struct *ssp) __acquires(ssp
> > > > >  	int retval;
> > > > >  
> > > > >  	if (IS_ENABLED(CONFIG_NEED_SRCU_NMI_SAFE))
> > > > > -		retval = __srcu_read_lock_nmisafe(ssp);
> > > > > +		retval = __srcu_read_lock_nmisafe(ssp, true);
> > > > >  	else
> > > > >  		retval = __srcu_read_lock(ssp);
> > > > 
> > > > Shouldn't it be checked also when CONFIG_NEED_SRCU_NMI_SAFE=n ?
> > > 
> > > You are asking why there is no "true" argument to __srcu_read_lock()?
> > > That is because it checks unconditionally.
> > 
> > It checks unconditionally but it always assumes not to be called as nmisafe.
> > 
> > For example on x86/arm64/loongarch, the same ssp used with both srcu_read_lock() and
> > srcu_read_lock_nmisafe() won't report an issue. But on powerpc it will.
> > 
> > My point is that strong archs should warn as well on behalf of others, to detect
> > mistakes early.
> 
> Good point, especially given that x86_64 and arm64 are a rather large
> fraction of the uses.  Not critically urgent, but definitely nice to have.

No indeed.

> 
> Did you by chance have a suggestion for a nice way to accomplish this?

This could be like this:

enum srcu_nmi_flags {
     SRCU_NMI_UNKNOWN = 0x0,
     SRCU_NMI_UNSAFE  = 0x1,
     SRCU_NMI_SAFE    = 0x2
};

#ifdef CONFIG_NEED_SRCU_NMI_SAFE
static inline int __srcu_read_lock_nmisafe(struct srcu_struct *ssp, enum srcu_nmi_flags flags)
{
	int idx;
	struct srcu_data *sdp = raw_cpu_ptr(ssp->sda);

	idx = READ_ONCE(ssp->srcu_idx) & 0x1;
	atomic_long_inc(&sdp->srcu_lock_count[idx]);
	smp_mb__after_atomic(); /* B */  /* Avoid leaking the critical section. */

	srcu_check_nmi_safety(ssp, flags);

	return idx;
}
#else
static inline int __srcu_read_lock_nmisafe(struct srcu_struct *ssp, enum srcu_nmi_flags flags)
{
	srcu_check_nmi_safety(ssp, flags);
	return __srcu_read_lock(ssp);
}
#endif

static inline int srcu_read_lock_nmisafe(struct srcu_struct *ssp)
{
	return  __srcu_read_lock_nmisafe(ssp, SRCU_NMI_SAFE);
}

// An __srcu_read_lock() caller in kernel/rcu/tasks.h must be
// taken care of as well
static inline int srcu_read_lock(struct srcu_struct *ssp)
{
	srcu_check_nmi_safety(ssp, SRCU_NMI_UNSAFE);
	return  __srcu_read_lock(ssp);
}

And then you can call __srcu_read_lock_nmisafe(ssp, SRCU_NMI_UNKNOWN) from
initializers of gp.