On Oct 24, 2022, at 9:01 PM, NeilBrown [off-list ref] wrote:

On Tue, 25 Oct 2022, Chuck Lever III wrote:

quoted

quoted

On Oct 24, 2022, at 7:43 PM, NeilBrown [off-list ref] wrote:

quoted

quoted

+	list = rhltable_lookup(&nfs4_file_rhltable, fhp,
+			       nfs4_file_rhash_params);
+	rhl_for_each_entry_rcu(fi, tmp, list, fi_rlist) {
		if (fh_match(&fi->fi_fhandle, &fhp->fh_handle)) {
			if (refcount_inc_not_zero(&fi->fi_ref))
				ret = fi;
		} else if (d_inode(fhp->fh_dentry) == fi->fi_inode)
			fi->fi_aliased = alias_found = true;

This d_inde)( == fi->fi_inode test is no longer relevant.  Everything in
the list must have the same inode.  Best remove it.

My understanding is that more than one inode can hash into one of
these bucket lists. I don't see how rhltable_insert() can prevent
that from happening, and that will certainly be true if we go back
to using a fixed-size table.

With rhltable each bucket is a list of lists.  An rhlist_head contains
two pointers.  "rhead.next" points to the next entry in the bucket which
will have a different key (different inode in this case).  "next" points
to the next entry with the same key - different filehandle, same inode,
in this case.

That's the detail I didn't yet have, thanks. I will remove the extra
inode test and give it a try.

The list you get back from rhltable_lookup() isn't the full bucket list.
It is just the list within the bucket of all elements with the same key.

--
Chuck Lever