Re: [PATCH] Btrfs: unlock extent when the dio completes for reads

[Josef Bacik <jbacik@fusionio.com>]

On Tue, Jul 31, 2012 at 03:30:49PM -0600, Zach Brown wrote:
> > +static void btrfs_direct_read_iodone(struct kiocb *iocb, loff_t offset,
> > +				     ssize_t bytes, void *private, int ret,
> > +				     bool is_async)
> > +{
> > +	struct inode *inode = iocb->ki_filp->f_mapping->host;
> > +	u64 lockend;
> > +
> > +	if (get_state_private(&BTRFS_I(inode)->io_tree, offset, &lockend)) {
> > +		WARN_ON(1);
> > +		goto out;
> > +	}
> > +
> > +	unlock_extent(&BTRFS_I(inode)->io_tree, offset, lockend);
> > +out:
> > +	if (is_async)
> > +		aio_complete(iocb, ret, 0);
> > +	inode_dio_done(inode);
> > +}
> 
> Hmm.  I'm worried that this can be called from interrupt context and the
> state/extent calls there seem to use bare spin locks.  Did you run this
> with lockdep?  Am I confused?
> 

Our read completions are run in a helper thread since we need to do things like
clear the extent state stuff and such so it's safe in this regard.  Course I hit
a lockup with xfstests 91 because there are some cases where we can return
errors without doing the dio complete (rightly so since we haven't setup the dio
yet), so now I have to figure out how to know when we've been unlocked via dio
complete or that I've gotten an error back and already unlocked it...

Josef