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

[Darrick J. Wong]

On Wed, Sep 06, 2023 at 08:01:56PM +0100, Matthew Wilcox wrote:
> On Wed, Sep 06, 2023 at 11:43:36AM -0700, Darrick J. Wong wrote:
> > On Wed, Sep 06, 2023 at 07:33:48PM +0100, Matthew Wilcox wrote:
> > > On Wed, Sep 06, 2023 at 10:07:24AM -0700, Darrick J. Wong wrote:
> > > > On Wed, Sep 06, 2023 at 05:00:14PM +0100, Matthew Wilcox wrote:
> > > > > +++ b/fs/xfs/xfs_inode.c
> > > > > @@ -361,7 +361,7 @@ xfs_isilocked(
> > > > >  {
> > > > >  	if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
> > > > >  		if (!(lock_flags & XFS_ILOCK_SHARED))
> > > > > -			return !!ip->i_lock.mr_writer;
> > > > > +			return rwsem_is_write_locked(&ip->i_lock.mr_lock);
> > > > 
> > > > You'd be better off converting this to:
> > > > 
> > > > 	return __xfs_rwsem_islocked(&ip->i_lock.mr_lock,
> > > > 				(lock_flags & XFS_ILOCK_SHARED));
> > > > 
> > > > And then fixing __xfs_rwsem_islocked to do:
> > > > 
> > > > static inline bool
> > > > __xfs_rwsem_islocked(
> > > > 	struct rw_semaphore	*rwsem,
> > > > 	bool			shared)
> > > > {
> > > > 	if (!debug_locks) {
> > > > 		if (!shared)
> > > > 			return rwsem_is_write_locked(rwsem);
> > > > 		return rwsem_is_locked(rwsem);
> > > > 	}
> > > > 
> > > > 	...
> > > > }
> > > 
> > > Thanks.
> > > 
> > > > > +++ b/include/linux/rwsem.h
> > > > > @@ -72,6 +72,11 @@ static inline int rwsem_is_locked(struct rw_semaphore *sem)
> > > > >  	return atomic_long_read(&sem->count) != 0;
> > > > >  }
> > > > >  
> > > > > +static inline int rwsem_is_write_locked(struct rw_semaphore *sem)
> > > > > +{
> > > > > +	return atomic_long_read(&sem->count) & 1;
> > > > 
> > > > 
> > > > atomic_long_read(&sem->count) & RWSEM_WRITER_LOCKED ?
> > > 
> > > Then this would either have to be in rwsem.c or we'd have to move the
> > > definition of RWSEM_WRITER_LOCKED to rwsem.h.  All three options are
> > > kind of bad.  I think I hate the bare '1' least.
> > 
> > I disagree, because using the bare 1 brings the most risk that someone
> > will subtly break the locking assertions some day when they get the
> > bright idea to move RWSEM_WRITER_LOCKED to the upper bit and fail to
> > notice this predicate and its magic number.  IMO moving it to the header
> > file (possibly with the usual __ prefix) would be preferable to leaving
> > a landmine.
> 
> +       return atomic_long_read(&sem->count) & 1 /* RWSEM_WRITER_LOCKED */;
> 
> works for you?

Yeah I guess that works.

--D