On Sat, May 8, 2021 at 7:28 PM Al Viro [off-list ref] wrote:

        Re READ_ONCE() - we are wrapped into
read_seqbegin_or_lock(&rename_lock, &seq) there, so it's more about
being explicit than about correctness considerations.

Well, part of this all is that the next step is that "vsnprintf()"
with '%pD' would basically use prepend_entries() with just the RCU
lock.

That said, even with the rename lock, that will only cause a retry on
rename - it won't necessarily fix any confusion that comes from the
compiler possibly silently re-loading 'parent' multiple times, and
getting different pointers due to a concurrent rename.

Now, those different results should all be individually ok, due to RCU
freeing, but it's _really_ confusing if 'parent' might be two
different things within the same iteration of the loop.

I don't see anything truly horrible that would happen - mainly "we'll
prefetch one parent, and then due to reloading the pointer we might
actually _use_ another parent entirely for the next iteration", but it
really is best to avoid that kind of confusion.

                Linus