[Intel-gfx] [PATCH 1/3] drm: use the lookup lock in drm_is_current_master

[Boqun Feng]

On Thu, Jul 22, 2021 at 09:02:41PM +0200, Daniel Vetter wrote:
> On Thu, Jul 22, 2021 at 6:00 PM Boqun Feng <boqun.feng at gmail.com> wrote:
> >
> > On Thu, Jul 22, 2021 at 12:38:10PM +0200, Daniel Vetter wrote:
> > > On Thu, Jul 22, 2021 at 05:29:27PM +0800, Desmond Cheong Zhi Xi wrote:
> > > > Inside drm_is_current_master, using the outer drm_device.master_mutex
> > > > to protect reads of drm_file.master makes the function prone to creating
> > > > lock hierarchy inversions. Instead, we can use the
> > > > drm_file.master_lookup_lock that sits at the bottom of the lock
> > > > hierarchy.
> > > >
> > > > Reported-by: Daniel Vetter <daniel.vetter at ffwll.ch>
> > > > Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx at gmail.com>
> > > > ---
> > > >  drivers/gpu/drm/drm_auth.c | 9 +++++----
> > > >  1 file changed, 5 insertions(+), 4 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/drm_auth.c b/drivers/gpu/drm/drm_auth.c
> > > > index f00354bec3fb..9c24b8cc8e36 100644
> > > > --- a/drivers/gpu/drm/drm_auth.c
> > > > +++ b/drivers/gpu/drm/drm_auth.c
> > > > @@ -63,8 +63,9 @@
> > > >
> > > >  static bool drm_is_current_master_locked(struct drm_file *fpriv)
> > > >  {
> > > > -   lockdep_assert_held_once(&fpriv->minor->dev->master_mutex);
> > > > -
> > > > +   /* Either drm_device.master_mutex or drm_file.master_lookup_lock
> > > > +    * should be held here.
> > > > +    */
> > >
> > > Disappointing that lockdep can't check or conditions for us, a
> > > lockdep_assert_held_either would be really neat in some cases.
> > >
> >
> > The implementation is not hard but I don't understand the usage, for
> > example, if we have a global variable x, and two locks L1 and L2, and
> > the function
> >
> >         void do_something_to_x(void)
> >         {
> >                 lockdep_assert_held_either(L1, L2);
> >                 x++;
> >         }
> >
> > and two call sites:
> >
> >         void f(void)
> >         {
> >                 lock(L1);
> >                 do_something_to_x();
> >                 unlock(L1);
> >         }
> >
> >         void g(void)
> >         {
> >                 lock(L2);
> >                 do_something_to_x();
> >                 unlock(L2);
> >         }
> >
> > , wouldn't it be racy if f() and g() called by two threads at the same
> > time? Usually I would expect there exists a third synchronazition
> > mechanism (say M), which synchronizes the calls to f() and g(), and we
> > put M in the lockdep_assert_held() check inside do_something_to_x()
> > like:
> >
> >         void do_something_to_x(void)
> >         {
> >                 lockdep_assert_held_once(M);
> >                 x++;
> >         }
> >
> > But of course, M may not be a lock, so we cannot put the assert there.
> >
> > My cscope failed to find ->master_lookup_lock in -rc2 and seems it's not
> > introduced in the patchset either, could you point me the branch this
> > patchset is based on, so that I could understand this better, and maybe
> > come up with a solution? Thanks ;-)
> 
> The use case is essentially 2 nesting locks, and only the innermost is
> used to update a field. So when you only read this field, it's safe if
> either of these two locks are held. Essentially this is a read/write lock
> type of thing, except for various reasons the two locks might not be of
> the same type (like here where the write lock is a mutex, but the read
> lock is a spinlock).
> 
> It's a bit like the rcu_derefence macro where it's ok to either be in a
> rcu_read_lock() section, or holding the relevant lock that's used to
> update the value. We do _not_ have two different locks that allow writing
> to the same X.
> 
> Does that make it clearer what's the use-case here?
> 
> In an example:
> 
> void * interesting_pointer.
> 
> do_update_interesting_pointer()
> {
> 	mutex_lock(A);
> 	/* do more stuff to prepare things */
> 	spin_lock(B);
> 	interesting_pointer = new_value;
> 	spin_unlock(B);
> 	mutex_unlock(A);
> }
> 
> read_interesting_thing_locked()
> {
> 	lockdep_assert_held_either(A, B);
> 
> 	return interesting_pointer->thing;
> }
> 
> read_interesting_thing()
> {
> 	int thing;
> 	spin_lock(B);
> 	thing = interesting_pointer->thing;
> 	spin_unlock(B);
> 
> 	return B;
> }
> 
> spinlock might also be irqsafe here if this can be called from irq
> context.
> 

Make sense, so we'd better also provide lockdep_assert_held_both(), I
think, to use it at the update side, something as below:


	/*
	 * lockdep_assert_held_{both,either}().
	 * 
	 * Sometimes users can use a combination of two locks to
	 * implement a rwlock-like lock, for example, say we have
	 * locks L1 and L2, and we only allow updates when two locks
	 * both held like:
	 * 
	 * update()
	 * {
	 *	lockdep_assert_held_both(L1, L2);
	 *	x++; // update x
	 * }
	 *
	 * while for read-only accesses, either lock suffices (since
	 * holding either lock means others cannot hold both, so readers
	 * serialized with the updaters):
	 *
	 * read()
	 * {
	 * 	lockdep_assert_held_either(L1, L2);
	 *	r = x; // read x
	 * }
	 */

	#define lockdep_assert_held_both(l1, l2)	do {			\
			WARN_ON_ONCE(debug_locks &&				\
					(!lockdep_is_held(l1) ||		\
					 !lockdep_is_held(l2)));		\
	} while (0)

	#define lockdep_assert_held_either(l1, l2)	do {			\
			WARN_ON_ONCE(debug_locks &&				\
					(!lockdep_is_held(l1) &&		\
					 !lockdep_is_held(l2)));		\
	} while (0)

Still need sometime to think through this (e.g. on whether this it the
best implementation).

Regards,
Boqun

> Cheers, Daniel
> 
> > Regards,
> > Boqun
> >
> > > Adding lockdep folks, maybe they have ideas.
> > >
> > > On the patch:
> > >
> > > Reviewed-by: Daniel Vetter <daniel.vetter at ffwll.ch>
> > >
> > > >     return fpriv->is_master && drm_lease_owner(fpriv->master) == fpriv->minor->dev->master;
> > > >  }
> > > >
> > > > @@ -82,9 +83,9 @@ bool drm_is_current_master(struct drm_file *fpriv)
> > > >  {
> > > >     bool ret;
> > > >
> > > > -   mutex_lock(&fpriv->minor->dev->master_mutex);
> > > > +   spin_lock(&fpriv->master_lookup_lock);
> > > >     ret = drm_is_current_master_locked(fpriv);
> > > > -   mutex_unlock(&fpriv->minor->dev->master_mutex);
> > > > +   spin_unlock(&fpriv->master_lookup_lock);
> > > >
> > > >     return ret;
> > > >  }
> > > > --
> > > > 2.25.1
> > > >
> > >
> > > --
> > > Daniel Vetter
> > > Software Engineer, Intel Corporation
> > > http://blog.ffwll.ch
> 
> 
> 
> -- 
> Daniel Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch