Re: [PATCH 08/14] xfs: support embedded dfops in transaction

[Brian Foster <bfoster@redhat.com>]

On Thu, Jul 19, 2018 at 12:56:53PM -0700, Christoph Hellwig wrote:
> > +/*
> > + * Move state from one xfs_defer_ops to another and reset the source to initial
> > + * state. This is primarily used to carry state forward across transaction rolls
> > + * with internal dfops.
> > + */
> 
> > +void
> > +xfs_defer_move(
> > +	struct xfs_defer_ops	*dst,
> > +	struct xfs_defer_ops	*src)
> > +{
> > +	int			i;
> > +
> > +	ASSERT(dst != src);
> > +
> > +	list_splice_init(&src->dop_intake, &dst->dop_intake);
> > +	list_splice_init(&src->dop_pending, &dst->dop_pending);
> > +
> > +	for (i = 0; i < XFS_DEFER_OPS_NR_INODES; i++) {
> > +		if (!src->dop_inodes[i])
> > +			break;
> > +		dst->dop_inodes[i] = src->dop_inodes[i];
> > +	}
> > +	for (i = 0; i< XFS_DEFER_OPS_NR_BUFS; i++) {
> > +		if (!src->dop_bufs[i])
> > +			break;
> > +		dst->dop_bufs[i] = src->dop_bufs[i];
> > +	}
> 
> I suspect simply memcpy()ing the state over would be faster..

Hm, I figured open-coding it would be faster than function calls since
there are only 2 pointers, but tbh I'm not sure either way and don't
have a strong preference. That said, I've been thinking a bit about how
to get rid of dop_bufs/inodes entirely (see my reply to the cover letter
thread). If that looks promising then it's probably not worth changing
this over since the code will go away (but otherwise I'll change it to
memcpy() if that's preferred)..

Brian