[PATCH] mm/mmu_notifier: rename mmu_notifier_synchronize() to <...>_barrier()

[PATCH] mm/mmu_notifier: rename mmu_notifier_synchronize() to <...>_barrier()

[Sean Christopherson]

...and update its comment to explicitly reference its association with
mmu_notifier_call_srcu().

Contrary to its name, mmu_notifier_synchronize() does not synchronize
the notifier's SRCU instance, but rather waits for RCU callbacks to
finished, i.e. it invokes rcu_barrier().  The RCU documentation is
quite clear on this matter, explicitly calling out that rcu_barrier()
does not imply synchronize_rcu().  The misnomer could lean an unwary
developer to incorrectly assume that mmu_notifier_synchronize() can
be used in conjunction with mmu_notifier_unregister_no_release() to
implement a variation of mmu_notifier_unregister() that synchronizes
SRCU without invoking ->release.  A Documentation-allergic and hasty
developer could be further confused by the fact that rcu_barrier() is
indeed a pass-through to synchronize_rcu()... in tiny SRCU.

Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
---
 mm/mmu_notifier.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/mm/mmu_notifier.c b/mm/mmu_notifier.c
index 5119ff846769..46ebea6483bf 100644
--- a/mm/mmu_notifier.c
+++ b/mm/mmu_notifier.c
@@ -35,12 +35,12 @@ void mmu_notifier_call_srcu(struct rcu_head *rcu,
 }
 EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
 
-void mmu_notifier_synchronize(void)
+void mmu_notifier_barrier(void)
 {
-	/* Wait for any running method to finish. */
+	/* Wait for any running RCU callbacks (see above) to finish. */
 	srcu_barrier(&srcu);
 }
-EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
+EXPORT_SYMBOL_GPL(mmu_notifier_barrier);
 
 /*
  * This function can't run concurrently against mmu_notifier_register
-- 
2.19.1

Re: [PATCH] mm/mmu_notifier: rename mmu_notifier_synchronize() to <...>_barrier()

[Andrew Morton]

On Mon,  5 Nov 2018 11:29:55 -0800 Sean Christopherson <sean.j.christopherson@intel.com> wrote:

> ...and update its comment to explicitly reference its association with
> mmu_notifier_call_srcu().
> 
> Contrary to its name, mmu_notifier_synchronize() does not synchronize
> the notifier's SRCU instance, but rather waits for RCU callbacks to
> finished, i.e. it invokes rcu_barrier().  The RCU documentation is
> quite clear on this matter, explicitly calling out that rcu_barrier()
> does not imply synchronize_rcu().  The misnomer could lean an unwary
> developer to incorrectly assume that mmu_notifier_synchronize() can
> be used in conjunction with mmu_notifier_unregister_no_release() to
> implement a variation of mmu_notifier_unregister() that synchronizes
> SRCU without invoking ->release.  A Documentation-allergic and hasty
> developer could be further confused by the fact that rcu_barrier() is
> indeed a pass-through to synchronize_rcu()... in tiny SRCU.

Fair enough.

> --- a/mm/mmu_notifier.c
> +++ b/mm/mmu_notifier.c
> @@ -35,12 +35,12 @@ void mmu_notifier_call_srcu(struct rcu_head *rcu,
>  }
>  EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
>  
> -void mmu_notifier_synchronize(void)
> +void mmu_notifier_barrier(void)
>  {
> -	/* Wait for any running method to finish. */
> +	/* Wait for any running RCU callbacks (see above) to finish. */
>  	srcu_barrier(&srcu);
>  }
> -EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
> +EXPORT_SYMBOL_GPL(mmu_notifier_barrier);
>  
>  /*
>   * This function can't run concurrently against mmu_notifier_register

But as it has no callers, why retain it?

Re: [PATCH] mm/mmu_notifier: rename mmu_notifier_synchronize() to <...>_barrier()

[Sean Christopherson]

On Mon, Nov 05, 2018 at 12:18:33PM -0800, Andrew Morton wrote:
> On Mon,  5 Nov 2018 11:29:55 -0800 Sean Christopherson <sean.j.christopherson@intel.com> wrote:
> 
> > ...and update its comment to explicitly reference its association with
> > mmu_notifier_call_srcu().
> > 
> > Contrary to its name, mmu_notifier_synchronize() does not synchronize
> > the notifier's SRCU instance, but rather waits for RCU callbacks to
> > finished, i.e. it invokes rcu_barrier().  The RCU documentation is
> > quite clear on this matter, explicitly calling out that rcu_barrier()
> > does not imply synchronize_rcu().  The misnomer could lean an unwary
> > developer to incorrectly assume that mmu_notifier_synchronize() can
> > be used in conjunction with mmu_notifier_unregister_no_release() to
> > implement a variation of mmu_notifier_unregister() that synchronizes
> > SRCU without invoking ->release.  A Documentation-allergic and hasty
> > developer could be further confused by the fact that rcu_barrier() is
> > indeed a pass-through to synchronize_rcu()... in tiny SRCU.
> 
> Fair enough.
> 
> > --- a/mm/mmu_notifier.c
> > +++ b/mm/mmu_notifier.c
> > @@ -35,12 +35,12 @@ void mmu_notifier_call_srcu(struct rcu_head *rcu,
> >  }
> >  EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
> >  
> > -void mmu_notifier_synchronize(void)
> > +void mmu_notifier_barrier(void)
> >  {
> > -	/* Wait for any running method to finish. */
> > +	/* Wait for any running RCU callbacks (see above) to finish. */
> >  	srcu_barrier(&srcu);
> >  }
> > -EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
> > +EXPORT_SYMBOL_GPL(mmu_notifier_barrier);
> >  
> >  /*
> >   * This function can't run concurrently against mmu_notifier_register
> 
> But as it has no callers, why retain it?

I was hesitant to remove it altogether since it was explicitly added to
complement mmu_notifier_call_srcu()[1] even though the initial user of
mmu_notifier_call_srcu() didn't use mmu_notifier_synchronize()[2].  I
assume there was a good reason for adding the barrier function, but
maybe that's a bad assumption.

[1] b972216e27d1 ("mmu_notifier: add call_srcu and sync function for listener to delay call and sync")
[2] https://lore.kernel.org/patchwork/patch/515318/

Re: [PATCH] mm/mmu_notifier: rename mmu_notifier_synchronize() to <...>_barrier()

[Matthew Wilcox]

On Mon, Nov 05, 2018 at 12:18:33PM -0800, Andrew Morton wrote:
> > +++ b/mm/mmu_notifier.c
> 
> But as it has no callers, why retain it?

... and this patch missed the declaration of mmu_notifier_synchronize
in include/linux/mmu_notifier.h (whether we delete it or rename it,
that mention of it needs to be fixed)

Re: [PATCH] mm/mmu_notifier: rename mmu_notifier_synchronize() to <...>_barrier()

[Sean Christopherson]

On Mon, 2018-11-05 at 13:14 -0800, Matthew Wilcox wrote:
> On Mon, Nov 05, 2018 at 12:18:33PM -0800, Andrew Morton wrote:
> > 
> > > 
> > > +++ b/mm/mmu_notifier.c
> > But as it has no callers, why retain it?
> ... and this patch missed the declaration of mmu_notifier_synchronize
> in include/linux/mmu_notifier.h (whether we delete it or rename it,
> that mention of it needs to be fixed)

Doh.  I'll remove the function and send v2.

Re: [PATCH] mm/mmu_notifier: rename mmu_notifier_synchronize() to <...>_barrier()

[Michal Hocko]

On Mon 05-11-18 12:18:33, Andrew Morton wrote:
> On Mon,  5 Nov 2018 11:29:55 -0800 Sean Christopherson <sean.j.christopherson@intel.com> wrote:
[...]
> > -EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
> > +EXPORT_SYMBOL_GPL(mmu_notifier_barrier);
> >  
> >  /*
> >   * This function can't run concurrently against mmu_notifier_register
> 
> But as it has no callers, why retain it?

Exported symbols are not freed and if this is not used by any in-kernel
code then I would just remove it.

-- 
Michal Hocko
SUSE Labs