On Wed, Mar 24, 2021 at 06:59:37PM +0100, Christoph Hellwig wrote:

On Wed, Mar 24, 2021 at 10:57:35AM -0700, Darrick J. Wong wrote:

quoted

On Wed, Mar 24, 2021 at 08:03:06AM +0100, Christoph Hellwig wrote:

quoted

Remove the generic xfs_inode_walk and just open code the only caller.

This is going in the wrong direction for me.  Maybe.

I was planning to combine the reclaim inode walk into this function, and
later on share it with inactivation.  This made for one switch-happy
iteration function, but it meant there was only one loop.

Ok, we can skip this for now if this gets in your way.  Or I can resend
a different patch 2 that just removes the no tag case for now.

quoted

OFC maybe the point that you and/or Dave were trying to make is that I
should be doing the opposite, and combining the inactivation loop into
what is now the (badly misnamed) xfs_reclaim_inodes_ag?  And leave this
blockgc loop alone?

That is my gut feeling.  No guarantee it actually works out, and given
that I've lead you down the wrong road a few times I already feel guily
ahead of time..

Actually, collapsing all of the tag walkers into xfs_inode_walk was
pretty straightforward, and in the end I just borrowed bits and pieces
from patches 2 and 3 to make it happen and clean up the arguments.  The
net change is 55 lines deleted and ~1k less code (granted with all the
debugging and ubsan crud turned on).

--D