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>

fs/nfs/Makefile

@@ -30,7 +30,7 @@ nfsv4-y := nfs4proc.o nfs4xdr.o nfs4state.o nfs4renewd.o nfs4super.o nfs4file.o
 nfsv4-$(CONFIG_NFS_USE_LEGACY_DNS) += cache_lib.o
 nfsv4-$(CONFIG_SYSCTL)	+= nfs4sysctl.o
 nfsv4-$(CONFIG_NFS_V4_1)	+= pnfs.o pnfs_dev.o pnfs_nfs.o
-nfsv4-$(CONFIG_NFS_V4_2)	+= nfs42proc.o
+nfsv4-$(CONFIG_NFS_V4_2)	+= nfs42proc.o nfs42xattr.o
 
 obj-$(CONFIG_PNFS_FILE_LAYOUT) += filelayout/
 obj-$(CONFIG_PNFS_BLOCK) += blocklayout/

fs/nfs/inode.c

@@ -193,6 +193,7 @@ bool nfs_check_cache_invalid(struct inode *inode, unsigned long flags)
 
 	return nfs_check_cache_invalid_not_delegated(inode, flags);
 }
+EXPORT_SYMBOL_GPL(nfs_check_cache_invalid);
 
 static void nfs_set_cache_invalid(struct inode *inode, unsigned long flags)
 {
@@ -234,11 +235,13 @@ static void nfs_zap_caches_locked(struct inode *inode)
 					| NFS_INO_INVALID_DATA
 					| NFS_INO_INVALID_ACCESS
 					| NFS_INO_INVALID_ACL
+					| NFS_INO_INVALID_XATTR
 					| NFS_INO_REVAL_PAGECACHE);
 	} else
 		nfs_set_cache_invalid(inode, NFS_INO_INVALID_ATTR
 					| NFS_INO_INVALID_ACCESS
 					| NFS_INO_INVALID_ACL
+					| NFS_INO_INVALID_XATTR
 					| NFS_INO_REVAL_PAGECACHE);
 	nfs_zap_label_cache_locked(nfsi);
 }
@@ -1897,7 +1900,8 @@ static int nfs_update_inode(struct inode *inode, struct nfs_fattr *fattr)
 			if (!(have_writers || have_delegation)) {
 				invalid |= NFS_INO_INVALID_DATA
 					| NFS_INO_INVALID_ACCESS
-					| NFS_INO_INVALID_ACL;
+					| NFS_INO_INVALID_ACL
+					| NFS_INO_INVALID_XATTR;
 				/* Force revalidate of all attributes */
 				save_cache_validity |= NFS_INO_INVALID_CTIME
 					| NFS_INO_INVALID_MTIME
@@ -2100,6 +2104,9 @@ struct inode *nfs_alloc_inode(struct super_block *sb)
 #if IS_ENABLED(CONFIG_NFS_V4)
 	nfsi->nfs4_acl = NULL;
 #endif /* CONFIG_NFS_V4 */
+#ifdef CONFIG_NFS_V4_2
+	nfsi->xattr_cache = NULL;
+#endif
 	return &nfsi->vfs_inode;
 }
 EXPORT_SYMBOL_GPL(nfs_alloc_inode);

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;

fs/nfs/nfs4_fs.h

@@ -626,12 +626,34 @@ static inline bool nfs4_state_match_open_stateid_other(const struct nfs4_state *
 		nfs4_stateid_match_other(&state->open_stateid, stateid);
 }
 
+/* nfs42xattr.c */
+#ifdef CONFIG_NFS_V4_2
+extern int __init nfs4_xattr_cache_init(void);
+extern void nfs4_xattr_cache_exit(void);
+extern void nfs4_xattr_cache_add(struct inode *inode, const char *name,
+				 const char *buf, struct page **pages,
+				 ssize_t buflen);
+extern void nfs4_xattr_cache_remove(struct inode *inode, const char *name);
+extern ssize_t nfs4_xattr_cache_get(struct inode *inode, const char *name,
+				char *buf, ssize_t buflen);
+extern void nfs4_xattr_cache_set_list(struct inode *inode, const char *buf,
+				      ssize_t buflen);
+extern ssize_t nfs4_xattr_cache_list(struct inode *inode, char *buf,
+				     ssize_t buflen);
+extern void nfs4_xattr_cache_zap(struct inode *inode);
 #else
+static inline void nfs4_xattr_cache_zap(struct inode *inode)
+{
+}
+#endif /* CONFIG_NFS_V4_2 */
+
+#else /* CONFIG_NFS_V4 */
 
 #define nfs4_close_state(a, b) do { } while (0)
 #define nfs4_close_sync(a, b) do { } while (0)
 #define nfs4_state_protect(a, b, c, d) do { } while (0)
 #define nfs4_state_protect_write(a, b, c, d) do { } while (0)
 
+
 #endif /* CONFIG_NFS_V4 */
 #endif /* __LINUX_FS_NFS_NFS4_FS.H */

fs/nfs/nfs4proc.c

@@ -7448,6 +7448,7 @@ static int nfs4_xattr_set_nfs4_user(const struct xattr_handler *handler,
 				    size_t buflen, int flags)
 {
 	struct nfs_access_entry cache;
+	int ret;
 
 	if (!nfs_server_capable(inode, NFS_CAP_XATTR))
 		return -EOPNOTSUPP;
@@ -7466,10 +7467,17 @@ static int nfs4_xattr_set_nfs4_user(const struct xattr_handler *handler,
 			return -EACCES;
 	}
 
-	if (buf == NULL)
-		return nfs42_proc_removexattr(inode, key);
-	else
-		return nfs42_proc_setxattr(inode, key, buf, buflen, flags);
+	if (buf == NULL) {
+		ret = nfs42_proc_removexattr(inode, key);
+		if (!ret)
+			nfs4_xattr_cache_remove(inode, key);
+	} else {
+		ret = nfs42_proc_setxattr(inode, key, buf, buflen, flags);
+		if (!ret)
+			nfs4_xattr_cache_add(inode, key, buf, NULL, buflen);
+	}
+
+	return ret;
 }
 
 static int nfs4_xattr_get_nfs4_user(const struct xattr_handler *handler,
@@ -7477,6 +7485,7 @@ static int nfs4_xattr_get_nfs4_user(const struct xattr_handler *handler,
 				    const char *key, void *buf, size_t buflen)
 {
 	struct nfs_access_entry cache;
+	ssize_t ret;
 
 	if (!nfs_server_capable(inode, NFS_CAP_XATTR))
 		return -EOPNOTSUPP;
@@ -7486,7 +7495,17 @@ static int nfs4_xattr_get_nfs4_user(const struct xattr_handler *handler,
 			return -EACCES;
 	}
 
-	return nfs42_proc_getxattr(inode, key, buf, buflen);
+	ret = nfs_revalidate_inode(NFS_SERVER(inode), inode);
+	if (ret)
+		return ret;
+
+	ret = nfs4_xattr_cache_get(inode, key, buf, buflen);
+	if (ret >= 0 || (ret < 0 && ret != -ENOENT))
+		return ret;
+
+	ret = nfs42_proc_getxattr(inode, key, buf, buflen);
+
+	return ret;
 }
 
 static ssize_t
@@ -7494,7 +7513,7 @@ nfs4_listxattr_nfs4_user(struct inode *inode, char *list, size_t list_len)
 {
 	u64 cookie;
 	bool eof;
-	int ret, size;
+	ssize_t ret, size;
 	char *buf;
 	size_t buflen;
 	struct nfs_access_entry cache;
@@ -7507,6 +7526,14 @@ nfs4_listxattr_nfs4_user(struct inode *inode, char *list, size_t list_len)
 			return 0;
 	}
 
+	ret = nfs_revalidate_inode(NFS_SERVER(inode), inode);
+	if (ret)
+		return ret;
+
+	ret = nfs4_xattr_cache_list(inode, list, list_len);
+	if (ret >= 0 || (ret < 0 && ret != -ENOENT))
+		return ret;
+
 	cookie = 0;
 	eof = false;
 	buflen = list_len ? list_len : XATTR_LIST_MAX;
@@ -7526,6 +7553,9 @@ nfs4_listxattr_nfs4_user(struct inode *inode, char *list, size_t list_len)
 		size += ret;
 	}
 
+	if (list_len)
+		nfs4_xattr_cache_set_list(inode, list, size);
+
 	return size;
 }
 

fs/nfs/nfs4super.c

@@ -69,6 +69,7 @@ static void nfs4_evict_inode(struct inode *inode)
 	pnfs_destroy_layout(NFS_I(inode));
 	/* First call standard NFS clear_inode() code */
 	nfs_clear_inode(inode);
+	nfs4_xattr_cache_zap(inode);
 }
 
 struct nfs_referral_count {
@@ -268,6 +269,12 @@ static int __init init_nfs_v4(void)
 	if (err)
 		goto out1;
 
+#ifdef CONFIG_NFS_V4_2
+	err = nfs4_xattr_cache_init();
+	if (err)
+		goto out2;
+#endif
+
 	err = nfs4_register_sysctl();
 	if (err)
 		goto out2;
@@ -288,6 +295,9 @@ static void __exit exit_nfs_v4(void)
 	nfs4_pnfs_v3_ds_connect_unload();
 
 	unregister_nfs_version(&nfs_v4);
+#ifdef CONFIG_NFS_V4_2
+	nfs4_xattr_cache_exit();
+#endif
 	nfs4_unregister_sysctl();
 	nfs_idmap_quit();
 	nfs_dns_resolver_destroy();

include/linux/nfs_fs.h

@@ -102,6 +102,8 @@ struct nfs_delegation;
 
 struct posix_acl;
 
+struct nfs4_xattr_cache;
+
 /*
  * nfs fs inode data in memory
  */
@@ -188,6 +190,10 @@ struct nfs_inode {
 	struct fscache_cookie	*fscache;
 #endif
 	struct inode		vfs_inode;
+
+#ifdef CONFIG_NFS_V4_2
+	struct nfs4_xattr_cache *xattr_cache;
+#endif
 };
 
 struct nfs4_copy_state {

include/uapi/linux/nfs_fs.h

@@ -56,6 +56,7 @@
 #define NFSDBG_PNFS		0x1000
 #define NFSDBG_PNFS_LD		0x2000
 #define NFSDBG_STATE		0x4000
+#define NFSDBG_XATTRCACHE	0x8000
 #define NFSDBG_ALL		0xFFFF