Re: [PATCH 4/4] NFS: use a hash table for delegation lookup

[Christoph Hellwig <hch@lst.de>]

On Mon, Jul 14, 2025 at 09:14:27AM -0400, Jeff Layton wrote:
> > +	delegation_buckets = roundup_pow_of_two(nfs_delegation_watermark / 16);
> > +	server->delegation_hash_mask = delegation_buckets - 1;
> > +	server->delegation_hash_table = kmalloc_array(delegation_buckets,
> > +			sizeof(*server->delegation_hash_table), GFP_KERNEL);
> > +	if (!server->delegation_hash_table)
> > +		goto out_free_server;
> > +	for (i = 0; i < delegation_buckets; i++)
> > +		INIT_HLIST_HEAD(&server->delegation_hash_table[i]);
> > +
> 
> This is going to get created for any mount, even v3 ones. It might be
> better to only bother with this for v4 mounts. Maybe do this in
> nfs4_server_common_setup() instead?

Yeah, good idea.

> Also, I wonder if you'd be better off using the rhashtable
> infrastructure instead of adding a fixed-size one? The number of
> delegations in flight is very workload-dependent, so a resizeable
> hashtable may be a better option.

I did try that first.  But the rhashtable expects the hash key to
be embedded into the structure that has the rhash_head embedded into
it, while delegations use the file handle embeded into the inode as
the key.  So we'd have to bloat the deleations with a key copy for
that.