Re: [RFC PATCH] vfs: add inode lockdep assertions

[Darrick J. Wong]

On Wed, Sep 06, 2023 at 05:23:42PM +0200, Bernd Schubert wrote:
> 
> 
> On 9/6/23 17:20, Matthew Wilcox wrote:
> > On Thu, Aug 31, 2023 at 05:14:14PM +0200, Mateusz Guzik wrote:
> > > +++ b/include/linux/fs.h
> > > @@ -842,6 +842,16 @@ static inline void inode_lock_shared_nested(struct inode *inode, unsigned subcla
> > >   	down_read_nested(&inode->i_rwsem, subclass);
> > >   }
> > > +static inline void inode_assert_locked(struct inode *inode)
> > > +{
> > > +	lockdep_assert_held(&inode->i_rwsem);
> > > +}
> > > +
> > > +static inline void inode_assert_write_locked(struct inode *inode)
> > > +{
> > > +	lockdep_assert_held_write(&inode->i_rwsem);
> > > +}
> > 
> > This mirrors what we have in mm, but it's only going to trigger on
> > builds that have lockdep enabled.  Lockdep is very expensive; it
> > easily doubles the time it takes to run xfstests on my laptop, so
> > I don't generally enable it.  So what we also have in MM is:
> > 
> > static inline void mmap_assert_write_locked(struct mm_struct *mm)
> > {
> >          lockdep_assert_held_write(&mm->mmap_lock);
> >          VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_lock), mm);
> > }
> > 
> > Now if you have lockdep enabled, you get the lockdep check which
> > gives you all the lovely lockdep information, but if you don't, you
> > at least get the cheap check that someone is holding the lock at all.
> > 
> > ie I would make this:
> > 
> > +static inline void inode_assert_write_locked(struct inode *inode)
> > +{
> > +	lockdep_assert_held_write(&inode->i_rwsem);
> > +	WARN_ON_ONCE(!inode_is_locked(inode));
> > +}
> > 
> > Maybe the locking people could give us a rwsem_is_write_locked()
> > predicate, but until then, this is the best solution we came up with.
> 
> 
> Which is exactly what I had suggested in the other thread :)

Or hoist the XFS mrlock, because it actually /does/ know if the rwsem is
held in shared or exclusive mode.

--D