UBIFS: improve znode splitting rules
authorAdrian Hunter <ext-adrian.hunter@nokia.com>
Fri, 5 Sep 2008 08:56:05 +0000 (11:56 +0300)
committerArtem Bityutskiy <Artem.Bityutskiy@nokia.com>
Tue, 30 Sep 2008 08:12:57 +0000 (11:12 +0300)
When inserting into a full znode it is split into two
znodes.  Because data node keys are usually consecutive,
it is better to try to keep them together.  This patch
does a better job of that.

Signed-off-by: Adrian Hunter <ext-adrian.hunter@nokia.com>
fs/ubifs/tnc.c

index 66dc571..e0878a4 100644 (file)
@@ -1962,7 +1962,7 @@ static int tnc_insert(struct ubifs_info *c, struct ubifs_znode *znode,
 {
        struct ubifs_znode *zn, *zi, *zp;
        int i, keep, move, appending = 0;
-       union ubifs_key *key = &zbr->key;
+       union ubifs_key *key = &zbr->key, *key1;
 
        ubifs_assert(n >= 0 && n <= c->fanout);
 
@@ -2003,20 +2003,33 @@ again:
        zn->level = znode->level;
 
        /* Decide where to split */
-       if (znode->level == 0 && n == c->fanout &&
-           key_type(c, key) == UBIFS_DATA_KEY) {
-               union ubifs_key *key1;
-
-               /*
-                * If this is an inode which is being appended - do not split
-                * it because no other zbranches can be inserted between
-                * zbranches of consecutive data nodes anyway.
-                */
-               key1 = &znode->zbranch[n - 1].key;
-               if (key_inum(c, key1) == key_inum(c, key) &&
-                   key_type(c, key1) == UBIFS_DATA_KEY &&
-                   key_block(c, key1) == key_block(c, key) - 1)
-                       appending = 1;
+       if (znode->level == 0 && key_type(c, key) == UBIFS_DATA_KEY) {
+               /* Try not to split consecutive data keys */
+               if (n == c->fanout) {
+                       key1 = &znode->zbranch[n - 1].key;
+                       if (key_inum(c, key1) == key_inum(c, key) &&
+                           key_type(c, key1) == UBIFS_DATA_KEY)
+                               appending = 1;
+               } else
+                       goto check_split;
+       } else if (appending && n != c->fanout) {
+               /* Try not to split consecutive data keys */
+               appending = 0;
+check_split:
+               if (n >= (c->fanout + 1) / 2) {
+                       key1 = &znode->zbranch[0].key;
+                       if (key_inum(c, key1) == key_inum(c, key) &&
+                           key_type(c, key1) == UBIFS_DATA_KEY) {
+                               key1 = &znode->zbranch[n].key;
+                               if (key_inum(c, key1) != key_inum(c, key) ||
+                                   key_type(c, key1) != UBIFS_DATA_KEY) {
+                                       keep = n;
+                                       move = c->fanout - keep;
+                                       zi = znode;
+                                       goto do_split;
+                               }
+                       }
+               }
        }
 
        if (appending) {
@@ -2046,6 +2059,8 @@ again:
                        zbr->znode->parent = zn;
        }
 
+do_split:
+
        __set_bit(DIRTY_ZNODE, &zn->flags);
        atomic_long_inc(&c->dirty_zn_cnt);
 
@@ -2072,14 +2087,11 @@ again:
 
        /* Insert new znode (produced by spitting) into the parent */
        if (zp) {
-               i = n;
+               if (n == 0 && zi == znode && znode->iip == 0)
+                       correct_parent_keys(c, znode);
+
                /* Locate insertion point */
                n = znode->iip + 1;
-               if (appending && n != c->fanout)
-                       appending = 0;
-
-               if (i == 0 && zi == znode && znode->iip == 0)
-                       correct_parent_keys(c, znode);
 
                /* Tail recursion */
                zbr->key = zn->zbranch[0].key;