diff --git a/fs/nfs/pnfs_nfs.c b/fs/nfs/pnfs_nfs.c
index fd2a2f0..3bb2b74 100644
--- a/fs/nfs/pnfs_nfs.c
+++ b/fs/nfs/pnfs_nfs.c
@@ -13,6 +13,8 @@
 #include "internal.h"
 #include "pnfs.h"
 
+#define NFSDBG_FACILITY		NFSDBG_PNFS
+
 static void pnfs_generic_fenceme(struct inode *inode,
 				 struct pnfs_layout_hdr *lo)
 {
@@ -290,3 +292,243 @@
 	return PNFS_ATTEMPTED;
 }
 EXPORT_SYMBOL_GPL(pnfs_generic_commit_pagelist);
+
+/*
+ * Data server cache
+ *
+ * Data servers can be mapped to different device ids.
+ * nfs4_pnfs_ds reference counting
+ *   - set to 1 on allocation
+ *   - incremented when a device id maps a data server already in the cache.
+ *   - decremented when deviceid is removed from the cache.
+ */
+static DEFINE_SPINLOCK(nfs4_ds_cache_lock);
+static LIST_HEAD(nfs4_data_server_cache);
+
+/* Debug routines */
+static void
+print_ds(struct nfs4_pnfs_ds *ds)
+{
+	if (ds == NULL) {
+		printk(KERN_WARNING "%s NULL device\n", __func__);
+		return;
+	}
+	printk(KERN_WARNING "        ds %s\n"
+		"        ref count %d\n"
+		"        client %p\n"
+		"        cl_exchange_flags %x\n",
+		ds->ds_remotestr,
+		atomic_read(&ds->ds_count), ds->ds_clp,
+		ds->ds_clp ? ds->ds_clp->cl_exchange_flags : 0);
+}
+
+static bool
+same_sockaddr(struct sockaddr *addr1, struct sockaddr *addr2)
+{
+	struct sockaddr_in *a, *b;
+	struct sockaddr_in6 *a6, *b6;
+
+	if (addr1->sa_family != addr2->sa_family)
+		return false;
+
+	switch (addr1->sa_family) {
+	case AF_INET:
+		a = (struct sockaddr_in *)addr1;
+		b = (struct sockaddr_in *)addr2;
+
+		if (a->sin_addr.s_addr == b->sin_addr.s_addr &&
+		    a->sin_port == b->sin_port)
+			return true;
+		break;
+
+	case AF_INET6:
+		a6 = (struct sockaddr_in6 *)addr1;
+		b6 = (struct sockaddr_in6 *)addr2;
+
+		/* LINKLOCAL addresses must have matching scope_id */
+		if (ipv6_addr_src_scope(&a6->sin6_addr) ==
+		    IPV6_ADDR_SCOPE_LINKLOCAL &&
+		    a6->sin6_scope_id != b6->sin6_scope_id)
+			return false;
+
+		if (ipv6_addr_equal(&a6->sin6_addr, &b6->sin6_addr) &&
+		    a6->sin6_port == b6->sin6_port)
+			return true;
+		break;
+
+	default:
+		dprintk("%s: unhandled address family: %u\n",
+			__func__, addr1->sa_family);
+		return false;
+	}
+
+	return false;
+}
+
+static bool
+_same_data_server_addrs_locked(const struct list_head *dsaddrs1,
+			       const struct list_head *dsaddrs2)
+{
+	struct nfs4_pnfs_ds_addr *da1, *da2;
+
+	/* step through both lists, comparing as we go */
+	for (da1 = list_first_entry(dsaddrs1, typeof(*da1), da_node),
+	     da2 = list_first_entry(dsaddrs2, typeof(*da2), da_node);
+	     da1 != NULL && da2 != NULL;
+	     da1 = list_entry(da1->da_node.next, typeof(*da1), da_node),
+	     da2 = list_entry(da2->da_node.next, typeof(*da2), da_node)) {
+		if (!same_sockaddr((struct sockaddr *)&da1->da_addr,
+				   (struct sockaddr *)&da2->da_addr))
+			return false;
+	}
+	if (da1 == NULL && da2 == NULL)
+		return true;
+
+	return false;
+}
+
+/*
+ * Lookup DS by addresses.  nfs4_ds_cache_lock is held
+ */
+static struct nfs4_pnfs_ds *
+_data_server_lookup_locked(const struct list_head *dsaddrs)
+{
+	struct nfs4_pnfs_ds *ds;
+
+	list_for_each_entry(ds, &nfs4_data_server_cache, ds_node)
+		if (_same_data_server_addrs_locked(&ds->ds_addrs, dsaddrs))
+			return ds;
+	return NULL;
+}
+
+static void destroy_ds(struct nfs4_pnfs_ds *ds)
+{
+	struct nfs4_pnfs_ds_addr *da;
+
+	dprintk("--> %s\n", __func__);
+	ifdebug(FACILITY)
+		print_ds(ds);
+
+	nfs_put_client(ds->ds_clp);
+
+	while (!list_empty(&ds->ds_addrs)) {
+		da = list_first_entry(&ds->ds_addrs,
+				      struct nfs4_pnfs_ds_addr,
+				      da_node);
+		list_del_init(&da->da_node);
+		kfree(da->da_remotestr);
+		kfree(da);
+	}
+
+	kfree(ds->ds_remotestr);
+	kfree(ds);
+}
+
+void nfs4_pnfs_ds_put(struct nfs4_pnfs_ds *ds)
+{
+	if (atomic_dec_and_lock(&ds->ds_count,
+				&nfs4_ds_cache_lock)) {
+		list_del_init(&ds->ds_node);
+		spin_unlock(&nfs4_ds_cache_lock);
+		destroy_ds(ds);
+	}
+}
+EXPORT_SYMBOL_GPL(nfs4_pnfs_ds_put);
+
+/*
+ * Create a string with a human readable address and port to avoid
+ * complicated setup around many dprinks.
+ */
+static char *
+nfs4_pnfs_remotestr(struct list_head *dsaddrs, gfp_t gfp_flags)
+{
+	struct nfs4_pnfs_ds_addr *da;
+	char *remotestr;
+	size_t len;
+	char *p;
+
+	len = 3;        /* '{', '}' and eol */
+	list_for_each_entry(da, dsaddrs, da_node) {
+		len += strlen(da->da_remotestr) + 1;    /* string plus comma */
+	}
+
+	remotestr = kzalloc(len, gfp_flags);
+	if (!remotestr)
+		return NULL;
+
+	p = remotestr;
+	*(p++) = '{';
+	len--;
+	list_for_each_entry(da, dsaddrs, da_node) {
+		size_t ll = strlen(da->da_remotestr);
+
+		if (ll > len)
+			goto out_err;
+
+		memcpy(p, da->da_remotestr, ll);
+		p += ll;
+		len -= ll;
+
+		if (len < 1)
+			goto out_err;
+		(*p++) = ',';
+		len--;
+	}
+	if (len < 2)
+		goto out_err;
+	*(p++) = '}';
+	*p = '\0';
+	return remotestr;
+out_err:
+	kfree(remotestr);
+	return NULL;
+}
+
+/*
+ * Given a list of multipath struct nfs4_pnfs_ds_addr, add it to ds cache if
+ * uncached and return cached struct nfs4_pnfs_ds.
+ */
+struct nfs4_pnfs_ds *
+nfs4_pnfs_ds_add(struct list_head *dsaddrs, gfp_t gfp_flags)
+{
+	struct nfs4_pnfs_ds *tmp_ds, *ds = NULL;
+	char *remotestr;
+
+	if (list_empty(dsaddrs)) {
+		dprintk("%s: no addresses defined\n", __func__);
+		goto out;
+	}
+
+	ds = kzalloc(sizeof(*ds), gfp_flags);
+	if (!ds)
+		goto out;
+
+	/* this is only used for debugging, so it's ok if its NULL */
+	remotestr = nfs4_pnfs_remotestr(dsaddrs, gfp_flags);
+
+	spin_lock(&nfs4_ds_cache_lock);
+	tmp_ds = _data_server_lookup_locked(dsaddrs);
+	if (tmp_ds == NULL) {
+		INIT_LIST_HEAD(&ds->ds_addrs);
+		list_splice_init(dsaddrs, &ds->ds_addrs);
+		ds->ds_remotestr = remotestr;
+		atomic_set(&ds->ds_count, 1);
+		INIT_LIST_HEAD(&ds->ds_node);
+		ds->ds_clp = NULL;
+		list_add(&ds->ds_node, &nfs4_data_server_cache);
+		dprintk("%s add new data server %s\n", __func__,
+			ds->ds_remotestr);
+	} else {
+		kfree(remotestr);
+		kfree(ds);
+		atomic_inc(&tmp_ds->ds_count);
+		dprintk("%s data server %s found, inc'ed ds_count to %d\n",
+			__func__, tmp_ds->ds_remotestr,
+			atomic_read(&tmp_ds->ds_count));
+		ds = tmp_ds;
+	}
+	spin_unlock(&nfs4_ds_cache_lock);
+out:
+	return ds;
+}
+EXPORT_SYMBOL_GPL(nfs4_pnfs_ds_add);