NFSv4.2: add client side xattr caching.

Implement client side caching for NFSv4.2 extended attributes. The cache
is a per-inode hashtable, with name/value entries. There is one special
entry for the listxattr cache.

NFS inodes have a pointer to a cache structure. The cache structure is
allocated on demand, freed when the cache is invalidated.

Memory shrinkers keep the size in check. Large entries (> PAGE_SIZE)
are collected by a separate shrinker, and freed more aggressively
than others.

Signed-off-by: Frank van der Linden <fllinden@amazon.com>
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>

1 files changed, 12 insertions, 0 deletions

diff --git a/fs/nfs/nfs42proc.c b/fs/nfs/nfs42proc.c
index 8c2e52bc986a..e200522469af 100644
--- a/fs/nfs/nfs42proc.c
+++ b/fs/nfs/nfs42proc.c
@@ -1182,6 +1182,18 @@ static ssize_t _nfs42_proc_getxattr(struct inode *inode, const char *name,
 	if (ret < 0)
 		return ret;
 
+	/*
+	 * Normally, the caching is done one layer up, but for successful
+	 * RPCS, always cache the result here, even if the caller was
+	 * just querying the length, or if the reply was too big for
+	 * the caller. This avoids a second RPC in the case of the
+	 * common query-alloc-retrieve cycle for xattrs.
+	 *
+	 * Note that xattr_len is always capped to XATTR_SIZE_MAX.
+	 */
+
+	nfs4_xattr_cache_add(inode, name, NULL, pages, res.xattr_len);
+
 	if (buflen) {
 		if (res.xattr_len > buflen)
 			return -ERANGE;