On Sat, May 8, 2021 at 2:06 PM Al Viro [off-list ref] wrote:

FWIW, if we go that way, I would make that

        while (dentry != root->dentry || &mnt->mnt != root->mnt) {
                int error;
                struct dentry *parent = READ_ONCE(dentry->d_parent);

Side note: you've added that READ_ONCE() to the parent reading, and I
think that's a bug-fix regardless. The old code does that plain

                parent = dentry->d_parent;

(after doing the mountpoint stuff). And d_parent isn't actually
guaranteed stable here.

It probably does not matter - we are in a RCU read-locked section, so
it's not like parent will go away, but in theory we might end up with
(for example) pre-fetching a different parent than the one we then
walk down.

But your READ_ONCE() is definitely the right thing to do (whether we
do your re-org or not, and whether we do this "prepend_buffer" thing
or not).

Do you want to do a final version with your fixes?

               Linus