On Wed, Jun 05, 2019 at 05:12:10PM +0200, Christoph Hellwig wrote:
> On Tue, Jun 04, 2019 at 04:24:42PM +1000, Dave Chinner wrote:
> > > @@ -1258,7 +1192,7 @@ xfs_buf_ioend_async(
> > > struct xfs_buf *bp)
> > > {
> > > INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
> > > - queue_work(bp->b_ioend_wq, &bp->b_ioend_work);
> > > + queue_work(bp->b_target->bt_mount->m_buf_workqueue, &bp->b_ioend_work);
> > > }
> >
> > It'd be nice to keep bp->b_ioend_wq to avoid pointer chasing here.
> > Perhaps we could set it up in _xfs_buf_alloc() where we are pretty
> > much guaranteed to have the xfs_mount hot in cache, and then it's
> > set for the life of the buffer. Just a thought, but either way:
>
> We'd need to benchmark this particular case, but the bloat to
> the xfs_buf structure probably cancels out at least one pointer
> dereference. That being said we probably could at least remove
> one indirection without bloating the buf, I'll take a look at that.
FWIW, I already plan to move the buf workqueue to the xfs_buftarg
so maybe that's the way to go here...
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
