ceph: add dir_layout to inode

Add a ceph_dir_layout to the inode, and calculate dentry hash values based on the parent directory's specified dir_hash function. This is needed because the old default Linux dcache hash function is extremely week and leads to a poor distribution of files among dir fragments. Signed-off-by: Sage Weil <sage@newdream.net>

ceph: add dir_layout to inode
Add a ceph_dir_layout to the inode, and calculate dentry hash values based on the parent directory's specified dir_hash function. This is needed because the old default Linux dcache hash function is extremely week and leads to a poor distribution of files among dir fragments. Signed-off-by: Sage Weil <sage@newdream.net>
Sage Weil
1 parent 3c0eee3fe6
Showing 6 changed files with 41 additions and 4 deletions Side-by-side Diff
fs/ceph/dir.c
fs/ceph/export.c
fs/ceph/inode.c
fs/ceph/super.h
include/linux/ceph/ceph_fs.h
net/ceph/ceph_hash.c
@@ -1216,6 +1216,26 @@
 	}
 }
  
+/*
+ * Return name hash for a given dentry.  This is dependent on
+ * the parent directory's hash function.
+ */
+unsigned ceph_dentry_hash(struct dentry *dn)
+{
+	struct inode *dir = dn->d_parent->d_inode;
+	struct ceph_inode_info *dci = ceph_inode(dir);
+
+	switch (dci->i_dir_layout.dl_dir_hash) {
+	case 0:	/* for backward compat */
+	case CEPH_STR_HASH_LINUX:
+		return dn->d_name.hash;
+
+	default:
+		return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
+				     dn->d_name.name, dn->d_name.len);
+	}
+}
+
 const struct file_operations ceph_dir_fops = {
 	.read = ceph_read_dir,
 	.readdir = ceph_readdir,
@@ -59,7 +59,7 @@
 		dout("encode_fh %p connectable\n", dentry);
 		cfh->ino = ceph_ino(dentry->d_inode);
 		cfh->parent_ino = ceph_ino(parent->d_inode);
-		cfh->parent_name_hash = parent->d_name.hash;
+		cfh->parent_name_hash = ceph_dentry_hash(parent);
 		*max_len = connected_handle_length;
 		type = 2;
 	} else if (*max_len >= handle_length) {
@@ -297,6 +297,8 @@
 	ci->i_release_count = 0;
 	ci->i_symlink = NULL;
  
+	memset(&ci->i_dir_layout, 0, sizeof(ci->i_dir_layout));
+
 	ci->i_fragtree = RB_ROOT;
 	mutex_init(&ci->i_fragtree_mutex);
  
@@ -239,6 +239,7 @@
 	unsigned i_ceph_flags;
 	unsigned long i_release_count;
  
+	struct ceph_dir_layout i_dir_layout;
 	struct ceph_file_layout i_layout;
 	char *i_symlink;
  
@@ -768,6 +769,7 @@
 extern void ceph_dentry_lru_touch(struct dentry *dn);
 extern void ceph_dentry_lru_del(struct dentry *dn);
 extern void ceph_invalidate_dentry_lease(struct dentry *dentry);
+extern unsigned ceph_dentry_hash(struct dentry *dn);
  
 /*
  * our d_ops vary depending on whether the inode is live,
@@ -43,6 +43,10 @@
 #define CEPH_FEATURE_NOSRCADDR      (1<<1)
 #define CEPH_FEATURE_MONCLOCKCHECK  (1<<2)
 #define CEPH_FEATURE_FLOCK          (1<<3)
+#define CEPH_FEATURE_SUBSCRIBE2     (1<<4)
+#define CEPH_FEATURE_MONNAMES       (1<<5)
+#define CEPH_FEATURE_RECONNECT_SEQ  (1<<6)
+#define CEPH_FEATURE_DIRLAYOUTHASH  (1<<7)
  
  
 /*
  
@@ -55,10 +59,10 @@
 	__le32 fl_stripe_count;    /* over this many objects */
 	__le32 fl_object_size;     /* until objects are this big, then move to
 				      new objects */
-	__le32 fl_cas_hash;        /* 0 = none; 1 = sha256 */
+	__le32 fl_cas_hash;        /* UNUSED.  0 = none; 1 = sha256 */
  
 	/* pg -> disk layout */
-	__le32 fl_object_stripe_unit;  /* for per-object parity, if any */
+	__le32 fl_object_stripe_unit;  /* UNUSED.  for per-object parity, if any */
  
 	/* object -> pg layout */
 	__le32 fl_pg_preferred; /* preferred primary for pg (-1 for none) */
@@ -69,6 +73,12 @@
  
 int ceph_file_layout_is_valid(const struct ceph_file_layout *layout);
  
+struct ceph_dir_layout {
+	__u8   dl_dir_hash;   /* see ceph_hash.h for ids */
+	__u8   dl_unused1;
+	__u16  dl_unused2;
+	__u32  dl_unused3;
+} __attribute__ ((packed));
  
 /* crypto algorithms */
 #define CEPH_CRYPTO_NONE 0x0
@@ -457,7 +467,7 @@
 	struct ceph_timespec rctime;
 	struct ceph_frag_tree_head fragtree;  /* (must be at end of struct) */
 } __attribute__ ((packed));
-/* followed by frag array, then symlink string, then xattr blob */
+/* followed by frag array, symlink string, dir layout, xattr blob */
  
 /* reply_lease follows dname, and reply_inode */
 struct ceph_mds_reply_lease {
  
 #include <linux/ceph/types.h>
+#include <linux/module.h>
  
 /*
  * Robert Jenkin's hash function.
@@ -104,6 +105,7 @@
 		return -1;
 	}
 }
+EXPORT_SYMBOL(ceph_str_hash);
  
 const char *ceph_str_hash_name(int type)
 {
@@ -116,4 +118,5 @@
 		return "unknown";
 	}
 }
+EXPORT_SYMBOL(ceph_str_hash_name);
...	...	@@ -1216,6 +1216,26 @@
1216	1216	}
1217	1217	}
1218	1218
	1219	+/*
	1220	+ * Return name hash for a given dentry. This is dependent on
	1221	+ * the parent directory's hash function.
	1222	+ */
	1223	+unsigned ceph_dentry_hash(struct dentry *dn)
	1224	+{
	1225	+ struct inode *dir = dn->d_parent->d_inode;
	1226	+ struct ceph_inode_info *dci = ceph_inode(dir);
	1227	+
	1228	+ switch (dci->i_dir_layout.dl_dir_hash) {
	1229	+ case 0: /* for backward compat */
	1230	+ case CEPH_STR_HASH_LINUX:
	1231	+ return dn->d_name.hash;
	1232	+
	1233	+ default:
	1234	+ return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
	1235	+ dn->d_name.name, dn->d_name.len);
	1236	+ }
	1237	+}
	1238	+
1219	1239	const struct file_operations ceph_dir_fops = {
1220	1240	.read = ceph_read_dir,
1221	1241	.readdir = ceph_readdir,
...	...	@@ -59,7 +59,7 @@
59	59	dout("encode_fh %p connectable\n", dentry);
60	60	cfh->ino = ceph_ino(dentry->d_inode);
61	61	cfh->parent_ino = ceph_ino(parent->d_inode);
62		- cfh->parent_name_hash = parent->d_name.hash;
	62	+ cfh->parent_name_hash = ceph_dentry_hash(parent);
63	63	*max_len = connected_handle_length;
64	64	type = 2;
65	65	} else if (*max_len >= handle_length) {
...	...	@@ -297,6 +297,8 @@
297	297	ci->i_release_count = 0;
298	298	ci->i_symlink = NULL;
299	299
	300	+ memset(&ci->i_dir_layout, 0, sizeof(ci->i_dir_layout));
	301	+
300	302	ci->i_fragtree = RB_ROOT;
301	303	mutex_init(&ci->i_fragtree_mutex);
302	304
...	...	@@ -239,6 +239,7 @@
239	239	unsigned i_ceph_flags;
240	240	unsigned long i_release_count;
241	241
	242	+ struct ceph_dir_layout i_dir_layout;
242	243	struct ceph_file_layout i_layout;
243	244	char *i_symlink;
244	245
...	...	@@ -768,6 +769,7 @@
768	769	extern void ceph_dentry_lru_touch(struct dentry *dn);
769	770	extern void ceph_dentry_lru_del(struct dentry *dn);
770	771	extern void ceph_invalidate_dentry_lease(struct dentry *dentry);
	772	+extern unsigned ceph_dentry_hash(struct dentry *dn);
771	773
772	774	/*
773	775	* our d_ops vary depending on whether the inode is live,
...	...	@@ -43,6 +43,10 @@
43	43	#define CEPH_FEATURE_NOSRCADDR (1<<1)
44	44	#define CEPH_FEATURE_MONCLOCKCHECK (1<<2)
45	45	#define CEPH_FEATURE_FLOCK (1<<3)
	46	+#define CEPH_FEATURE_SUBSCRIBE2 (1<<4)
	47	+#define CEPH_FEATURE_MONNAMES (1<<5)
	48	+#define CEPH_FEATURE_RECONNECT_SEQ (1<<6)
	49	+#define CEPH_FEATURE_DIRLAYOUTHASH (1<<7)
46	50
47	51
48	52	/*
49	53
...	...	@@ -55,10 +59,10 @@
55	59	__le32 fl_stripe_count; /* over this many objects */
56	60	__le32 fl_object_size; /* until objects are this big, then move to
57	61	new objects */
58		- __le32 fl_cas_hash; /* 0 = none; 1 = sha256 */
	62	+ __le32 fl_cas_hash; /* UNUSED. 0 = none; 1 = sha256 */
59	63
60	64	/* pg -> disk layout */
61		- __le32 fl_object_stripe_unit; /* for per-object parity, if any */
	65	+ __le32 fl_object_stripe_unit; /* UNUSED. for per-object parity, if any */
62	66
63	67	/* object -> pg layout */
64	68	__le32 fl_pg_preferred; /* preferred primary for pg (-1 for none) */
...	...	@@ -69,6 +73,12 @@
69	73
70	74	int ceph_file_layout_is_valid(const struct ceph_file_layout *layout);
71	75
	76	+struct ceph_dir_layout {
	77	+ __u8 dl_dir_hash; /* see ceph_hash.h for ids */
	78	+ __u8 dl_unused1;
	79	+ __u16 dl_unused2;
	80	+ __u32 dl_unused3;
	81	+} __attribute__ ((packed));
72	82
73	83	/* crypto algorithms */
74	84	#define CEPH_CRYPTO_NONE 0x0
...	...	@@ -457,7 +467,7 @@
457	467	struct ceph_timespec rctime;
458	468	struct ceph_frag_tree_head fragtree; /* (must be at end of struct) */
459	469	} __attribute__ ((packed));
460		-/* followed by frag array, then symlink string, then xattr blob */
	470	+/* followed by frag array, symlink string, dir layout, xattr blob */
461	471
462	472	/* reply_lease follows dname, and reply_inode */
463	473	struct ceph_mds_reply_lease {
1	1
2	2	#include <linux/ceph/types.h>
	3	+#include <linux/module.h>
3	4
4	5	/*
5	6	* Robert Jenkin's hash function.
...	...	@@ -104,6 +105,7 @@
104	105	return -1;
105	106	}
106	107	}
	108	+EXPORT_SYMBOL(ceph_str_hash);
107	109
108	110	const char *ceph_str_hash_name(int type)
109	111	{
...	...	@@ -116,4 +118,5 @@
116	118	return "unknown";
117	119	}
118	120	}
	121	+EXPORT_SYMBOL(ceph_str_hash_name);