[PATCH] ext3: add extent map support
authorAlex Tomas <alex@clusterfs.com>
Wed, 11 Oct 2006 08:21:03 +0000 (01:21 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Wed, 11 Oct 2006 18:14:16 +0000 (11:14 -0700)
On disk extents format:
/*
* this is extent on-disk structure
* it's used at the bottom of the tree
*/
struct ext3_extent {
__le32  ee_block;       /* first logical block extent covers */
__le16  ee_len;         /* number of blocks covered by extent */
__le16  ee_start_hi;    /* high 16 bits of physical block */
__le32  ee_start;       /* low 32 bigs of physical block */
};

Signed-off-by: Alex Tomas <alex@clusterfs.com>
Signed-off-by: Dave Kleikamp <shaggy@austin.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
12 files changed:
fs/ext4/Makefile
fs/ext4/dir.c
fs/ext4/extents.c [new file with mode: 0644]
fs/ext4/ialloc.c
fs/ext4/inode.c
fs/ext4/ioctl.c
fs/ext4/super.c
include/linux/ext4_fs.h
include/linux/ext4_fs_extents.h [new file with mode: 0644]
include/linux/ext4_fs_i.h
include/linux/ext4_fs_sb.h
include/linux/ext4_jbd2.h

index 09c487893e4aa8f2c1d01bff2c53a50518b71c3e..a6acb96ebeb9b004324632cb867412850f75938d 100644 (file)
@@ -5,7 +5,7 @@
 obj-$(CONFIG_EXT4DEV_FS) += ext4dev.o
 
 ext4dev-y      := balloc.o bitmap.o dir.o file.o fsync.o ialloc.o inode.o \
-          ioctl.o namei.o super.o symlink.o hash.o resize.o
+          ioctl.o namei.o super.o symlink.o hash.o resize.o extents.o
 
 ext4dev-$(CONFIG_EXT4DEV_FS_XATTR)     += xattr.o xattr_user.o xattr_trusted.o
 ext4dev-$(CONFIG_EXT4DEV_FS_POSIX_ACL) += acl.o
index 9833d5d00c46c308c1432361d4d71eb3ef0f2e82..18ac173af575c82d14e7fdd04d0eac114b761e96 100644 (file)
@@ -134,8 +134,7 @@ static int ext4_readdir(struct file * filp,
                struct buffer_head *bh = NULL;
 
                map_bh.b_state = 0;
-               err = ext4_get_blocks_handle(NULL, inode, blk, 1,
-                                               &map_bh, 0, 0);
+               err = ext4_get_blocks_wrap(NULL, inode, blk, 1, &map_bh, 0, 0);
                if (err > 0) {
                        page_cache_readahead(sb->s_bdev->bd_inode->i_mapping,
                                &filp->f_ra,
diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
new file mode 100644 (file)
index 0000000..f67b2ef
--- /dev/null
@@ -0,0 +1,2075 @@
+/*
+ * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
+ * Written by Alex Tomas <alex@clusterfs.com>
+ *
+ * Architecture independence:
+ *   Copyright (c) 2005, Bull S.A.
+ *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public Licens
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
+ */
+
+/*
+ * Extents support for EXT4
+ *
+ * TODO:
+ *   - ext4*_error() should be used in some situations
+ *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
+ *   - smart tree reduction
+ */
+
+#include <linux/module.h>
+#include <linux/fs.h>
+#include <linux/time.h>
+#include <linux/ext4_jbd2.h>
+#include <linux/jbd.h>
+#include <linux/smp_lock.h>
+#include <linux/highuid.h>
+#include <linux/pagemap.h>
+#include <linux/quotaops.h>
+#include <linux/string.h>
+#include <linux/slab.h>
+#include <linux/ext4_fs_extents.h>
+#include <asm/uaccess.h>
+
+
+static int ext4_ext_check_header(const char *function, struct inode *inode,
+                               struct ext4_extent_header *eh)
+{
+       const char *error_msg = NULL;
+
+       if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
+               error_msg = "invalid magic";
+               goto corrupted;
+       }
+       if (unlikely(eh->eh_max == 0)) {
+               error_msg = "invalid eh_max";
+               goto corrupted;
+       }
+       if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
+               error_msg = "invalid eh_entries";
+               goto corrupted;
+       }
+       return 0;
+
+corrupted:
+       ext4_error(inode->i_sb, function,
+                       "bad header in inode #%lu: %s - magic %x, "
+                       "entries %u, max %u, depth %u",
+                       inode->i_ino, error_msg, le16_to_cpu(eh->eh_magic),
+                       le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
+                       le16_to_cpu(eh->eh_depth));
+
+       return -EIO;
+}
+
+static handle_t *ext4_ext_journal_restart(handle_t *handle, int needed)
+{
+       int err;
+
+       if (handle->h_buffer_credits > needed)
+               return handle;
+       if (!ext4_journal_extend(handle, needed))
+               return handle;
+       err = ext4_journal_restart(handle, needed);
+
+       return handle;
+}
+
+/*
+ * could return:
+ *  - EROFS
+ *  - ENOMEM
+ */
+static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
+                               struct ext4_ext_path *path)
+{
+       if (path->p_bh) {
+               /* path points to block */
+               return ext4_journal_get_write_access(handle, path->p_bh);
+       }
+       /* path points to leaf/index in inode body */
+       /* we use in-core data, no need to protect them */
+       return 0;
+}
+
+/*
+ * could return:
+ *  - EROFS
+ *  - ENOMEM
+ *  - EIO
+ */
+static int ext4_ext_dirty(handle_t *handle, struct inode *inode,
+                               struct ext4_ext_path *path)
+{
+       int err;
+       if (path->p_bh) {
+               /* path points to block */
+               err = ext4_journal_dirty_metadata(handle, path->p_bh);
+       } else {
+               /* path points to leaf/index in inode body */
+               err = ext4_mark_inode_dirty(handle, inode);
+       }
+       return err;
+}
+
+static int ext4_ext_find_goal(struct inode *inode,
+                             struct ext4_ext_path *path,
+                             unsigned long block)
+{
+       struct ext4_inode_info *ei = EXT4_I(inode);
+       unsigned long bg_start;
+       unsigned long colour;
+       int depth;
+
+       if (path) {
+               struct ext4_extent *ex;
+               depth = path->p_depth;
+
+               /* try to predict block placement */
+               if ((ex = path[depth].p_ext))
+                       return le32_to_cpu(ex->ee_start)
+                                       + (block - le32_to_cpu(ex->ee_block));
+
+               /* it looks index is empty
+                * try to find starting from index itself */
+               if (path[depth].p_bh)
+                       return path[depth].p_bh->b_blocknr;
+       }
+
+       /* OK. use inode's group */
+       bg_start = (ei->i_block_group * EXT4_BLOCKS_PER_GROUP(inode->i_sb)) +
+               le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_first_data_block);
+       colour = (current->pid % 16) *
+                       (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
+       return bg_start + colour + block;
+}
+
+static int
+ext4_ext_new_block(handle_t *handle, struct inode *inode,
+                       struct ext4_ext_path *path,
+                       struct ext4_extent *ex, int *err)
+{
+       int goal, newblock;
+
+       goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
+       newblock = ext4_new_block(handle, inode, goal, err);
+       return newblock;
+}
+
+static inline int ext4_ext_space_block(struct inode *inode)
+{
+       int size;
+
+       size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
+                       / sizeof(struct ext4_extent);
+#ifdef AGRESSIVE_TEST
+       if (size > 6)
+               size = 6;
+#endif
+       return size;
+}
+
+static inline int ext4_ext_space_block_idx(struct inode *inode)
+{
+       int size;
+
+       size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
+                       / sizeof(struct ext4_extent_idx);
+#ifdef AGRESSIVE_TEST
+       if (size > 5)
+               size = 5;
+#endif
+       return size;
+}
+
+static inline int ext4_ext_space_root(struct inode *inode)
+{
+       int size;
+
+       size = sizeof(EXT4_I(inode)->i_data);
+       size -= sizeof(struct ext4_extent_header);
+       size /= sizeof(struct ext4_extent);
+#ifdef AGRESSIVE_TEST
+       if (size > 3)
+               size = 3;
+#endif
+       return size;
+}
+
+static inline int ext4_ext_space_root_idx(struct inode *inode)
+{
+       int size;
+
+       size = sizeof(EXT4_I(inode)->i_data);
+       size -= sizeof(struct ext4_extent_header);
+       size /= sizeof(struct ext4_extent_idx);
+#ifdef AGRESSIVE_TEST
+       if (size > 4)
+               size = 4;
+#endif
+       return size;
+}
+
+#ifdef EXT_DEBUG
+static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
+{
+       int k, l = path->p_depth;
+
+       ext_debug("path:");
+       for (k = 0; k <= l; k++, path++) {
+               if (path->p_idx) {
+                 ext_debug("  %d->%d", le32_to_cpu(path->p_idx->ei_block),
+                           le32_to_cpu(path->p_idx->ei_leaf));
+               } else if (path->p_ext) {
+                       ext_debug("  %d:%d:%d",
+                                 le32_to_cpu(path->p_ext->ee_block),
+                                 le16_to_cpu(path->p_ext->ee_len),
+                                 le32_to_cpu(path->p_ext->ee_start));
+               } else
+                       ext_debug("  []");
+       }
+       ext_debug("\n");
+}
+
+static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
+{
+       int depth = ext_depth(inode);
+       struct ext4_extent_header *eh;
+       struct ext4_extent *ex;
+       int i;
+
+       if (!path)
+               return;
+
+       eh = path[depth].p_hdr;
+       ex = EXT_FIRST_EXTENT(eh);
+
+       for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
+               ext_debug("%d:%d:%d ", le32_to_cpu(ex->ee_block),
+                         le16_to_cpu(ex->ee_len),
+                         le32_to_cpu(ex->ee_start));
+       }
+       ext_debug("\n");
+}
+#else
+#define ext4_ext_show_path(inode,path)
+#define ext4_ext_show_leaf(inode,path)
+#endif
+
+static void ext4_ext_drop_refs(struct ext4_ext_path *path)
+{
+       int depth = path->p_depth;
+       int i;
+
+       for (i = 0; i <= depth; i++, path++)
+               if (path->p_bh) {
+                       brelse(path->p_bh);
+                       path->p_bh = NULL;
+               }
+}
+
+/*
+ * binary search for closest index by given block
+ */
+static void
+ext4_ext_binsearch_idx(struct inode *inode, struct ext4_ext_path *path, int block)
+{
+       struct ext4_extent_header *eh = path->p_hdr;
+       struct ext4_extent_idx *r, *l, *m;
+
+       BUG_ON(eh->eh_magic != EXT4_EXT_MAGIC);
+       BUG_ON(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max));
+       BUG_ON(le16_to_cpu(eh->eh_entries) <= 0);
+
+       ext_debug("binsearch for %d(idx):  ", block);
+
+       l = EXT_FIRST_INDEX(eh) + 1;
+       r = EXT_FIRST_INDEX(eh) + le16_to_cpu(eh->eh_entries) - 1;
+       while (l <= r) {
+               m = l + (r - l) / 2;
+               if (block < le32_to_cpu(m->ei_block))
+                       r = m - 1;
+               else
+                       l = m + 1;
+               ext_debug("%p(%u):%p(%u):%p(%u) ", l, l->ei_block,
+                               m, m->ei_block, r, r->ei_block);
+       }
+
+       path->p_idx = l - 1;
+       ext_debug("  -> %d->%d ", le32_to_cpu(path->p_idx->ei_block),
+                 le32_to_cpu(path->p_idx->ei_leaf));
+
+#ifdef CHECK_BINSEARCH
+       {
+               struct ext4_extent_idx *chix, *ix;
+               int k;
+
+               chix = ix = EXT_FIRST_INDEX(eh);
+               for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
+                 if (k != 0 &&
+                     le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
+                               printk("k=%d, ix=0x%p, first=0x%p\n", k,
+                                       ix, EXT_FIRST_INDEX(eh));
+                               printk("%u <= %u\n",
+                                      le32_to_cpu(ix->ei_block),
+                                      le32_to_cpu(ix[-1].ei_block));
+                       }
+                       BUG_ON(k && le32_to_cpu(ix->ei_block)
+                                          <= le32_to_cpu(ix[-1].ei_block));
+                       if (block < le32_to_cpu(ix->ei_block))
+                               break;
+                       chix = ix;
+               }
+               BUG_ON(chix != path->p_idx);
+       }
+#endif
+
+}
+
+/*
+ * binary search for closest extent by given block
+ */
+static void
+ext4_ext_binsearch(struct inode *inode, struct ext4_ext_path *path, int block)
+{
+       struct ext4_extent_header *eh = path->p_hdr;
+       struct ext4_extent *r, *l, *m;
+
+       BUG_ON(eh->eh_magic != EXT4_EXT_MAGIC);
+       BUG_ON(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max));
+
+       if (eh->eh_entries == 0) {
+               /*
+                * this leaf is empty yet:
+                *  we get such a leaf in split/add case
+                */
+               return;
+       }
+
+       ext_debug("binsearch for %d:  ", block);
+
+       l = EXT_FIRST_EXTENT(eh) + 1;
+       r = EXT_FIRST_EXTENT(eh) + le16_to_cpu(eh->eh_entries) - 1;
+
+       while (l <= r) {
+               m = l + (r - l) / 2;
+               if (block < le32_to_cpu(m->ee_block))
+                       r = m - 1;
+               else
+                       l = m + 1;
+               ext_debug("%p(%u):%p(%u):%p(%u) ", l, l->ee_block,
+                               m, m->ee_block, r, r->ee_block);
+       }
+
+       path->p_ext = l - 1;
+       ext_debug("  -> %d:%d:%d ",
+                       le32_to_cpu(path->p_ext->ee_block),
+                       le32_to_cpu(path->p_ext->ee_start),
+                       le16_to_cpu(path->p_ext->ee_len));
+
+#ifdef CHECK_BINSEARCH
+       {
+               struct ext4_extent *chex, *ex;
+               int k;
+
+               chex = ex = EXT_FIRST_EXTENT(eh);
+               for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
+                       BUG_ON(k && le32_to_cpu(ex->ee_block)
+                                         <= le32_to_cpu(ex[-1].ee_block));
+                       if (block < le32_to_cpu(ex->ee_block))
+                               break;
+                       chex = ex;
+               }
+               BUG_ON(chex != path->p_ext);
+       }
+#endif
+
+}
+
+int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
+{
+       struct ext4_extent_header *eh;
+
+       eh = ext_inode_hdr(inode);
+       eh->eh_depth = 0;
+       eh->eh_entries = 0;
+       eh->eh_magic = EXT4_EXT_MAGIC;
+       eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode));
+       ext4_mark_inode_dirty(handle, inode);
+       ext4_ext_invalidate_cache(inode);
+       return 0;
+}
+
+struct ext4_ext_path *
+ext4_ext_find_extent(struct inode *inode, int block, struct ext4_ext_path *path)
+{
+       struct ext4_extent_header *eh;
+       struct buffer_head *bh;
+       short int depth, i, ppos = 0, alloc = 0;
+
+       eh = ext_inode_hdr(inode);
+       BUG_ON(eh == NULL);
+       if (ext4_ext_check_header(__FUNCTION__, inode, eh))
+               return ERR_PTR(-EIO);
+
+       i = depth = ext_depth(inode);
+
+       /* account possible depth increase */
+       if (!path) {
+               path = kmalloc(sizeof(struct ext4_ext_path) * (depth + 2),
+                               GFP_NOFS);
+               if (!path)
+                       return ERR_PTR(-ENOMEM);
+               alloc = 1;
+       }
+       memset(path, 0, sizeof(struct ext4_ext_path) * (depth + 1));
+       path[0].p_hdr = eh;
+
+       /* walk through the tree */
+       while (i) {
+               ext_debug("depth %d: num %d, max %d\n",
+                         ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
+               ext4_ext_binsearch_idx(inode, path + ppos, block);
+               path[ppos].p_block = le32_to_cpu(path[ppos].p_idx->ei_leaf);
+               path[ppos].p_depth = i;
+               path[ppos].p_ext = NULL;
+
+               bh = sb_bread(inode->i_sb, path[ppos].p_block);
+               if (!bh)
+                       goto err;
+
+               eh = ext_block_hdr(bh);
+               ppos++;
+               BUG_ON(ppos > depth);
+               path[ppos].p_bh = bh;
+               path[ppos].p_hdr = eh;
+               i--;
+
+               if (ext4_ext_check_header(__FUNCTION__, inode, eh))
+                       goto err;
+       }
+
+       path[ppos].p_depth = i;
+       path[ppos].p_hdr = eh;
+       path[ppos].p_ext = NULL;
+       path[ppos].p_idx = NULL;
+
+       if (ext4_ext_check_header(__FUNCTION__, inode, eh))
+               goto err;
+
+       /* find extent */
+       ext4_ext_binsearch(inode, path + ppos, block);
+
+       ext4_ext_show_path(inode, path);
+
+       return path;
+
+err:
+       ext4_ext_drop_refs(path);
+       if (alloc)
+               kfree(path);
+       return ERR_PTR(-EIO);
+}
+
+/*
+ * insert new index [logical;ptr] into the block at cupr
+ * it check where to insert: before curp or after curp
+ */
+static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
+                               struct ext4_ext_path *curp,
+                               int logical, int ptr)
+{
+       struct ext4_extent_idx *ix;
+       int len, err;
+
+       if ((err = ext4_ext_get_access(handle, inode, curp)))
+               return err;
+
+       BUG_ON(logical == le32_to_cpu(curp->p_idx->ei_block));
+       len = EXT_MAX_INDEX(curp->p_hdr) - curp->p_idx;
+       if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
+               /* insert after */
+               if (curp->p_idx != EXT_LAST_INDEX(curp->p_hdr)) {
+                       len = (len - 1) * sizeof(struct ext4_extent_idx);
+                       len = len < 0 ? 0 : len;
+                       ext_debug("insert new index %d after: %d. "
+                                       "move %d from 0x%p to 0x%p\n",
+                                       logical, ptr, len,
+                                       (curp->p_idx + 1), (curp->p_idx + 2));
+                       memmove(curp->p_idx + 2, curp->p_idx + 1, len);
+               }
+               ix = curp->p_idx + 1;
+       } else {
+               /* insert before */
+               len = len * sizeof(struct ext4_extent_idx);
+               len = len < 0 ? 0 : len;
+               ext_debug("insert new index %d before: %d. "
+                               "move %d from 0x%p to 0x%p\n",
+                               logical, ptr, len,
+                               curp->p_idx, (curp->p_idx + 1));
+               memmove(curp->p_idx + 1, curp->p_idx, len);
+               ix = curp->p_idx;
+       }
+
+       ix->ei_block = cpu_to_le32(logical);
+       ix->ei_leaf = cpu_to_le32(ptr);
+       curp->p_hdr->eh_entries = cpu_to_le16(le16_to_cpu(curp->p_hdr->eh_entries)+1);
+
+       BUG_ON(le16_to_cpu(curp->p_hdr->eh_entries)
+                            > le16_to_cpu(curp->p_hdr->eh_max));
+       BUG_ON(ix > EXT_LAST_INDEX(curp->p_hdr));
+
+       err = ext4_ext_dirty(handle, inode, curp);
+       ext4_std_error(inode->i_sb, err);
+
+       return err;
+}
+
+/*
+ * routine inserts new subtree into the path, using free index entry
+ * at depth 'at:
+ *  - allocates all needed blocks (new leaf and all intermediate index blocks)
+ *  - makes decision where to split
+ *  - moves remaining extens and index entries (right to the split point)
+ *    into the newly allocated blocks
+ *  - initialize subtree
+ */
+static int ext4_ext_split(handle_t *handle, struct inode *inode,
+                               struct ext4_ext_path *path,
+                               struct ext4_extent *newext, int at)
+{
+       struct buffer_head *bh = NULL;
+       int depth = ext_depth(inode);
+       struct ext4_extent_header *neh;
+       struct ext4_extent_idx *fidx;
+       struct ext4_extent *ex;
+       int i = at, k, m, a;
+       unsigned long newblock, oldblock;
+       __le32 border;
+       int *ablocks = NULL; /* array of allocated blocks */
+       int err = 0;
+
+       /* make decision: where to split? */
+       /* FIXME: now desicion is simplest: at current extent */
+
+       /* if current leaf will be splitted, then we should use
+        * border from split point */
+       BUG_ON(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr));
+       if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
+               border = path[depth].p_ext[1].ee_block;
+               ext_debug("leaf will be splitted."
+                               " next leaf starts at %d\n",
+                                 le32_to_cpu(border));
+       } else {
+               border = newext->ee_block;
+               ext_debug("leaf will be added."
+                               " next leaf starts at %d\n",
+                               le32_to_cpu(border));
+       }
+
+       /*
+        * if error occurs, then we break processing
+        * and turn filesystem read-only. so, index won't
+        * be inserted and tree will be in consistent
+        * state. next mount will repair buffers too
+        */
+
+       /*
+        * get array to track all allocated blocks
+        * we need this to handle errors and free blocks
+        * upon them
+        */
+       ablocks = kmalloc(sizeof(unsigned long) * depth, GFP_NOFS);
+       if (!ablocks)
+               return -ENOMEM;
+       memset(ablocks, 0, sizeof(unsigned long) * depth);
+
+       /* allocate all needed blocks */
+       ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
+       for (a = 0; a < depth - at; a++) {
+               newblock = ext4_ext_new_block(handle, inode, path, newext, &err);
+               if (newblock == 0)
+                       goto cleanup;
+               ablocks[a] = newblock;
+       }
+
+       /* initialize new leaf */
+       newblock = ablocks[--a];
+       BUG_ON(newblock == 0);
+       bh = sb_getblk(inode->i_sb, newblock);
+       if (!bh) {
+               err = -EIO;
+               goto cleanup;
+       }
+       lock_buffer(bh);
+
+       if ((err = ext4_journal_get_create_access(handle, bh)))
+               goto cleanup;
+
+       neh = ext_block_hdr(bh);
+       neh->eh_entries = 0;
+       neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode));
+       neh->eh_magic = EXT4_EXT_MAGIC;
+       neh->eh_depth = 0;
+       ex = EXT_FIRST_EXTENT(neh);
+
+       /* move remain of path[depth] to the new leaf */
+       BUG_ON(path[depth].p_hdr->eh_entries != path[depth].p_hdr->eh_max);
+       /* start copy from next extent */
+       /* TODO: we could do it by single memmove */
+       m = 0;
+       path[depth].p_ext++;
+       while (path[depth].p_ext <=
+                       EXT_MAX_EXTENT(path[depth].p_hdr)) {
+               ext_debug("move %d:%d:%d in new leaf %lu\n",
+                               le32_to_cpu(path[depth].p_ext->ee_block),
+                               le32_to_cpu(path[depth].p_ext->ee_start),
+                               le16_to_cpu(path[depth].p_ext->ee_len),
+                               newblock);
+               /*memmove(ex++, path[depth].p_ext++,
+                               sizeof(struct ext4_extent));
+               neh->eh_entries++;*/
+               path[depth].p_ext++;
+               m++;
+       }
+       if (m) {
+               memmove(ex, path[depth].p_ext-m, sizeof(struct ext4_extent)*m);
+               neh->eh_entries = cpu_to_le16(le16_to_cpu(neh->eh_entries)+m);
+       }
+
+       set_buffer_uptodate(bh);
+       unlock_buffer(bh);
+
+       if ((err = ext4_journal_dirty_metadata(handle, bh)))
+               goto cleanup;
+       brelse(bh);
+       bh = NULL;
+
+       /* correct old leaf */
+       if (m) {
+               if ((err = ext4_ext_get_access(handle, inode, path + depth)))
+                       goto cleanup;
+               path[depth].p_hdr->eh_entries =
+                    cpu_to_le16(le16_to_cpu(path[depth].p_hdr->eh_entries)-m);
+               if ((err = ext4_ext_dirty(handle, inode, path + depth)))
+                       goto cleanup;
+
+       }
+
+       /* create intermediate indexes */
+       k = depth - at - 1;
+       BUG_ON(k < 0);
+       if (k)
+               ext_debug("create %d intermediate indices\n", k);
+       /* insert new index into current index block */
+       /* current depth stored in i var */
+       i = depth - 1;
+       while (k--) {
+               oldblock = newblock;
+               newblock = ablocks[--a];
+               bh = sb_getblk(inode->i_sb, newblock);
+               if (!bh) {
+                       err = -EIO;
+                       goto cleanup;
+               }
+               lock_buffer(bh);
+
+               if ((err = ext4_journal_get_create_access(handle, bh)))
+                       goto cleanup;
+
+               neh = ext_block_hdr(bh);
+               neh->eh_entries = cpu_to_le16(1);
+               neh->eh_magic = EXT4_EXT_MAGIC;
+               neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode));
+               neh->eh_depth = cpu_to_le16(depth - i);
+               fidx = EXT_FIRST_INDEX(neh);
+               fidx->ei_block = border;
+               fidx->ei_leaf = cpu_to_le32(oldblock);
+
+               ext_debug("int.index at %d (block %lu): %lu -> %lu\n", i,
+                               newblock, (unsigned long) le32_to_cpu(border),
+                               oldblock);
+               /* copy indexes */
+               m = 0;
+               path[i].p_idx++;
+
+               ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
+                               EXT_MAX_INDEX(path[i].p_hdr));
+               BUG_ON(EXT_MAX_INDEX(path[i].p_hdr) !=
+                               EXT_LAST_INDEX(path[i].p_hdr));
+               while (path[i].p_idx <= EXT_MAX_INDEX(path[i].p_hdr)) {
+                       ext_debug("%d: move %d:%d in new index %lu\n", i,
+                                       le32_to_cpu(path[i].p_idx->ei_block),
+                                       le32_to_cpu(path[i].p_idx->ei_leaf),
+                                       newblock);
+                       /*memmove(++fidx, path[i].p_idx++,
+                                       sizeof(struct ext4_extent_idx));
+                       neh->eh_entries++;
+                       BUG_ON(neh->eh_entries > neh->eh_max);*/
+                       path[i].p_idx++;
+                       m++;
+               }
+               if (m) {
+                       memmove(++fidx, path[i].p_idx - m,
+                               sizeof(struct ext4_extent_idx) * m);
+                       neh->eh_entries =
+                               cpu_to_le16(le16_to_cpu(neh->eh_entries) + m);
+               }
+               set_buffer_uptodate(bh);
+               unlock_buffer(bh);
+
+               if ((err = ext4_journal_dirty_metadata(handle, bh)))
+                       goto cleanup;
+               brelse(bh);
+               bh = NULL;
+
+               /* correct old index */
+               if (m) {
+                       err = ext4_ext_get_access(handle, inode, path + i);
+                       if (err)
+                               goto cleanup;
+                       path[i].p_hdr->eh_entries = cpu_to_le16(le16_to_cpu(path[i].p_hdr->eh_entries)-m);
+                       err = ext4_ext_dirty(handle, inode, path + i);
+                       if (err)
+                               goto cleanup;
+               }
+
+               i--;
+       }
+
+       /* insert new index */
+       if (err)
+               goto cleanup;
+
+       err = ext4_ext_insert_index(handle, inode, path + at,
+                                   le32_to_cpu(border), newblock);
+
+cleanup:
+       if (bh) {
+               if (buffer_locked(bh))
+                       unlock_buffer(bh);
+               brelse(bh);
+       }
+
+       if (err) {
+               /* free all allocated blocks in error case */
+               for (i = 0; i < depth; i++) {
+                       if (!ablocks[i])
+                               continue;
+                       ext4_free_blocks(handle, inode, ablocks[i], 1);
+               }
+       }
+       kfree(ablocks);
+
+       return err;
+}
+
+/*
+ * routine implements tree growing procedure:
+ *  - allocates new block
+ *  - moves top-level data (index block or leaf) into the new block
+ *  - initialize new top-level, creating index that points to the
+ *    just created block
+ */
+static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
+                                       struct ext4_ext_path *path,
+                                       struct ext4_extent *newext)
+{
+       struct ext4_ext_path *curp = path;
+       struct ext4_extent_header *neh;
+       struct ext4_extent_idx *fidx;
+       struct buffer_head *bh;
+       unsigned long newblock;
+       int err = 0;
+
+       newblock = ext4_ext_new_block(handle, inode, path, newext, &err);
+       if (newblock == 0)
+               return err;
+
+       bh = sb_getblk(inode->i_sb, newblock);
+       if (!bh) {
+               err = -EIO;
+               ext4_std_error(inode->i_sb, err);
+               return err;
+       }
+       lock_buffer(bh);
+
+       if ((err = ext4_journal_get_create_access(handle, bh))) {
+               unlock_buffer(bh);
+               goto out;
+       }
+
+       /* move top-level index/leaf into new block */
+       memmove(bh->b_data, curp->p_hdr, sizeof(EXT4_I(inode)->i_data));
+
+       /* set size of new block */
+       neh = ext_block_hdr(bh);
+       /* old root could have indexes or leaves
+        * so calculate e_max right way */
+       if (ext_depth(inode))
+         neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode));
+       else
+         neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode));
+       neh->eh_magic = EXT4_EXT_MAGIC;
+       set_buffer_uptodate(bh);
+       unlock_buffer(bh);
+
+       if ((err = ext4_journal_dirty_metadata(handle, bh)))
+               goto out;
+
+       /* create index in new top-level index: num,max,pointer */
+       if ((err = ext4_ext_get_access(handle, inode, curp)))
+               goto out;
+
+       curp->p_hdr->eh_magic = EXT4_EXT_MAGIC;
+       curp->p_hdr->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode));
+       curp->p_hdr->eh_entries = cpu_to_le16(1);
+       curp->p_idx = EXT_FIRST_INDEX(curp->p_hdr);
+       /* FIXME: it works, but actually path[0] can be index */
+       curp->p_idx->ei_block = EXT_FIRST_EXTENT(path[0].p_hdr)->ee_block;
+       curp->p_idx->ei_leaf = cpu_to_le32(newblock);
+
+       neh = ext_inode_hdr(inode);
+       fidx = EXT_FIRST_INDEX(neh);
+       ext_debug("new root: num %d(%d), lblock %d, ptr %d\n",
+                 le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
+                 le32_to_cpu(fidx->ei_block), le32_to_cpu(fidx->ei_leaf));
+
+       neh->eh_depth = cpu_to_le16(path->p_depth + 1);
+       err = ext4_ext_dirty(handle, inode, curp);
+out:
+       brelse(bh);
+
+       return err;
+}
+
+/*
+ * routine finds empty index and adds new leaf. if no free index found
+ * then it requests in-depth growing
+ */
+static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
+                                       struct ext4_ext_path *path,
+                                       struct ext4_extent *newext)
+{
+       struct ext4_ext_path *curp;
+       int depth, i, err = 0;
+
+repeat:
+       i = depth = ext_depth(inode);
+
+       /* walk up to the tree and look for free index entry */
+       curp = path + depth;
+       while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
+               i--;
+               curp--;
+       }
+
+       /* we use already allocated block for index block
+        * so, subsequent data blocks should be contigoues */
+       if (EXT_HAS_FREE_INDEX(curp)) {
+               /* if we found index with free entry, then use that
+                * entry: create all needed subtree and add new leaf */
+               err = ext4_ext_split(handle, inode, path, newext, i);
+
+               /* refill path */
+               ext4_ext_drop_refs(path);
+               path = ext4_ext_find_extent(inode,
+                                           le32_to_cpu(newext->ee_block),
+                                           path);
+               if (IS_ERR(path))
+                       err = PTR_ERR(path);
+       } else {
+               /* tree is full, time to grow in depth */
+               err = ext4_ext_grow_indepth(handle, inode, path, newext);
+               if (err)
+                       goto out;
+
+               /* refill path */
+               ext4_ext_drop_refs(path);
+               path = ext4_ext_find_extent(inode,
+                                           le32_to_cpu(newext->ee_block),
+                                           path);
+               if (IS_ERR(path)) {
+                       err = PTR_ERR(path);
+                       goto out;
+               }
+
+               /*
+                * only first (depth 0 -> 1) produces free space
+                * in all other cases we have to split growed tree
+                */
+               depth = ext_depth(inode);
+               if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
+                       /* now we need split */
+                       goto repeat;
+               }
+       }
+
+out:
+       return err;
+}
+
+/*
+ * returns allocated block in subsequent extent or EXT_MAX_BLOCK
+ * NOTE: it consider block number from index entry as
+ * allocated block. thus, index entries have to be consistent
+ * with leafs
+ */
+static unsigned long
+ext4_ext_next_allocated_block(struct ext4_ext_path *path)
+{
+       int depth;
+
+       BUG_ON(path == NULL);
+       depth = path->p_depth;
+
+       if (depth == 0 && path->p_ext == NULL)
+               return EXT_MAX_BLOCK;
+
+       while (depth >= 0) {
+               if (depth == path->p_depth) {
+                       /* leaf */
+                       if (path[depth].p_ext !=
+                                       EXT_LAST_EXTENT(path[depth].p_hdr))
+                         return le32_to_cpu(path[depth].p_ext[1].ee_block);
+               } else {
+                       /* index */
+                       if (path[depth].p_idx !=
+                                       EXT_LAST_INDEX(path[depth].p_hdr))
+                         return le32_to_cpu(path[depth].p_idx[1].ei_block);
+               }
+               depth--;
+       }
+
+       return EXT_MAX_BLOCK;
+}
+
+/*
+ * returns first allocated block from next leaf or EXT_MAX_BLOCK
+ */
+static unsigned ext4_ext_next_leaf_block(struct inode *inode,
+                                               struct ext4_ext_path *path)
+{
+       int depth;
+
+       BUG_ON(path == NULL);
+       depth = path->p_depth;
+
+       /* zero-tree has no leaf blocks at all */
+       if (depth == 0)
+               return EXT_MAX_BLOCK;
+
+       /* go to index block */
+       depth--;
+
+       while (depth >= 0) {
+               if (path[depth].p_idx !=
+                               EXT_LAST_INDEX(path[depth].p_hdr))
+                 return le32_to_cpu(path[depth].p_idx[1].ei_block);
+               depth--;
+       }
+
+       return EXT_MAX_BLOCK;
+}
+
+/*
+ * if leaf gets modified and modified extent is first in the leaf
+ * then we have to correct all indexes above
+ * TODO: do we need to correct tree in all cases?
+ */
+int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
+                               struct ext4_ext_path *path)
+{
+       struct ext4_extent_header *eh;
+       int depth = ext_depth(inode);
+       struct ext4_extent *ex;
+       __le32 border;
+       int k, err = 0;
+
+       eh = path[depth].p_hdr;
+       ex = path[depth].p_ext;
+       BUG_ON(ex == NULL);
+       BUG_ON(eh == NULL);
+
+       if (depth == 0) {
+               /* there is no tree at all */
+               return 0;
+       }
+
+       if (ex != EXT_FIRST_EXTENT(eh)) {
+               /* we correct tree if first leaf got modified only */
+               return 0;
+       }
+
+       /*
+        * TODO: we need correction if border is smaller then current one
+        */
+       k = depth - 1;
+       border = path[depth].p_ext->ee_block;
+       if ((err = ext4_ext_get_access(handle, inode, path + k)))
+               return err;
+       path[k].p_idx->ei_block = border;
+       if ((err = ext4_ext_dirty(handle, inode, path + k)))
+               return err;
+
+       while (k--) {
+               /* change all left-side indexes */
+               if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
+                       break;
+               if ((err = ext4_ext_get_access(handle, inode, path + k)))
+                       break;
+               path[k].p_idx->ei_block = border;
+               if ((err = ext4_ext_dirty(handle, inode, path + k)))
+                       break;
+       }
+
+       return err;
+}
+
+static int inline
+ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
+                               struct ext4_extent *ex2)
+{
+       /* FIXME: 48bit support */
+        if (le32_to_cpu(ex1->ee_block) + le16_to_cpu(ex1->ee_len)
+           != le32_to_cpu(ex2->ee_block))
+               return 0;
+
+#ifdef AGRESSIVE_TEST
+       if (le16_to_cpu(ex1->ee_len) >= 4)
+               return 0;
+#endif
+
+        if (le32_to_cpu(ex1->ee_start) + le16_to_cpu(ex1->ee_len)
+                       == le32_to_cpu(ex2->ee_start))
+               return 1;
+       return 0;
+}
+
+/*
+ * this routine tries to merge requsted extent into the existing
+ * extent or inserts requested extent as new one into the tree,
+ * creating new leaf in no-space case
+ */
+int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
+                               struct ext4_ext_path *path,
+                               struct ext4_extent *newext)
+{
+       struct ext4_extent_header * eh;
+       struct ext4_extent *ex, *fex;
+       struct ext4_extent *nearex; /* nearest extent */
+       struct ext4_ext_path *npath = NULL;
+       int depth, len, err, next;
+
+       BUG_ON(newext->ee_len == 0);
+       depth = ext_depth(inode);
+       ex = path[depth].p_ext;
+       BUG_ON(path[depth].p_hdr == NULL);
+
+       /* try to insert block into found extent and return */
+       if (ex && ext4_can_extents_be_merged(inode, ex, newext)) {
+               ext_debug("append %d block to %d:%d (from %d)\n",
+                               le16_to_cpu(newext->ee_len),
+                               le32_to_cpu(ex->ee_block),
+                               le16_to_cpu(ex->ee_len),
+                               le32_to_cpu(ex->ee_start));
+               if ((err = ext4_ext_get_access(handle, inode, path + depth)))
+                       return err;
+               ex->ee_len = cpu_to_le16(le16_to_cpu(ex->ee_len)
+                                        + le16_to_cpu(newext->ee_len));
+               eh = path[depth].p_hdr;
+               nearex = ex;
+               goto merge;
+       }
+
+repeat:
+       depth = ext_depth(inode);
+       eh = path[depth].p_hdr;
+       if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
+               goto has_space;
+
+       /* probably next leaf has space for us? */
+       fex = EXT_LAST_EXTENT(eh);
+       next = ext4_ext_next_leaf_block(inode, path);
+       if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block)
+           && next != EXT_MAX_BLOCK) {
+               ext_debug("next leaf block - %d\n", next);
+               BUG_ON(npath != NULL);
+               npath = ext4_ext_find_extent(inode, next, NULL);
+               if (IS_ERR(npath))
+                       return PTR_ERR(npath);
+               BUG_ON(npath->p_depth != path->p_depth);
+               eh = npath[depth].p_hdr;
+               if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
+                       ext_debug("next leaf isnt full(%d)\n",
+                                 le16_to_cpu(eh->eh_entries));
+                       path = npath;
+                       goto repeat;
+               }
+               ext_debug("next leaf has no free space(%d,%d)\n",
+                         le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
+       }
+
+       /*
+        * there is no free space in found leaf
+        * we're gonna add new leaf in the tree
+        */
+       err = ext4_ext_create_new_leaf(handle, inode, path, newext);
+       if (err)
+               goto cleanup;
+       depth = ext_depth(inode);
+       eh = path[depth].p_hdr;
+
+has_space:
+       nearex = path[depth].p_ext;
+
+       if ((err = ext4_ext_get_access(handle, inode, path + depth)))
+               goto cleanup;
+
+       if (!nearex) {
+               /* there is no extent in this leaf, create first one */
+               ext_debug("first extent in the leaf: %d:%d:%d\n",
+                               le32_to_cpu(newext->ee_block),
+                               le32_to_cpu(newext->ee_start),
+                               le16_to_cpu(newext->ee_len));
+               path[depth].p_ext = EXT_FIRST_EXTENT(eh);
+       } else if (le32_to_cpu(newext->ee_block)
+                          > le32_to_cpu(nearex->ee_block)) {
+/*             BUG_ON(newext->ee_block == nearex->ee_block); */
+               if (nearex != EXT_LAST_EXTENT(eh)) {
+                       len = EXT_MAX_EXTENT(eh) - nearex;
+                       len = (len - 1) * sizeof(struct ext4_extent);
+                       len = len < 0 ? 0 : len;
+                       ext_debug("insert %d:%d:%d after: nearest 0x%p, "
+                                       "move %d from 0x%p to 0x%p\n",
+                                       le32_to_cpu(newext->ee_block),
+                                       le32_to_cpu(newext->ee_start),
+                                       le16_to_cpu(newext->ee_len),
+                                       nearex, len, nearex + 1, nearex + 2);
+                       memmove(nearex + 2, nearex + 1, len);
+               }
+               path[depth].p_ext = nearex + 1;
+       } else {
+               BUG_ON(newext->ee_block == nearex->ee_block);
+               len = (EXT_MAX_EXTENT(eh) - nearex) * sizeof(struct ext4_extent);
+               len = len < 0 ? 0 : len;
+               ext_debug("insert %d:%d:%d before: nearest 0x%p, "
+                               "move %d from 0x%p to 0x%p\n",
+                               le32_to_cpu(newext->ee_block),
+                               le32_to_cpu(newext->ee_start),
+                               le16_to_cpu(newext->ee_len),
+                               nearex, len, nearex + 1, nearex + 2);
+               memmove(nearex + 1, nearex, len);
+               path[depth].p_ext = nearex;
+       }
+
+       eh->eh_entries = cpu_to_le16(le16_to_cpu(eh->eh_entries)+1);
+       nearex = path[depth].p_ext;
+       nearex->ee_block = newext->ee_block;
+       nearex->ee_start = newext->ee_start;
+       nearex->ee_len = newext->ee_len;
+       /* FIXME: support for large fs */
+       nearex->ee_start_hi = 0;
+
+merge:
+       /* try to merge extents to the right */
+       while (nearex < EXT_LAST_EXTENT(eh)) {
+               if (!ext4_can_extents_be_merged(inode, nearex, nearex + 1))
+                       break;
+               /* merge with next extent! */
+               nearex->ee_len = cpu_to_le16(le16_to_cpu(nearex->ee_len)
+                                            + le16_to_cpu(nearex[1].ee_len));
+               if (nearex + 1 < EXT_LAST_EXTENT(eh)) {
+                       len = (EXT_LAST_EXTENT(eh) - nearex - 1)
+                                       * sizeof(struct ext4_extent);
+                       memmove(nearex + 1, nearex + 2, len);
+               }
+               eh->eh_entries = cpu_to_le16(le16_to_cpu(eh->eh_entries)-1);
+               BUG_ON(eh->eh_entries == 0);
+       }
+
+       /* try to merge extents to the left */
+
+       /* time to correct all indexes above */
+       err = ext4_ext_correct_indexes(handle, inode, path);
+       if (err)
+               goto cleanup;
+
+       err = ext4_ext_dirty(handle, inode, path + depth);
+
+cleanup:
+       if (npath) {
+               ext4_ext_drop_refs(npath);
+               kfree(npath);
+       }
+       ext4_ext_tree_changed(inode);
+       ext4_ext_invalidate_cache(inode);
+       return err;
+}
+
+int ext4_ext_walk_space(struct inode *inode, unsigned long block,
+                       unsigned long num, ext_prepare_callback func,
+                       void *cbdata)
+{
+       struct ext4_ext_path *path = NULL;
+       struct ext4_ext_cache cbex;
+       struct ext4_extent *ex;
+       unsigned long next, start = 0, end = 0;
+       unsigned long last = block + num;
+       int depth, exists, err = 0;
+
+       BUG_ON(func == NULL);
+       BUG_ON(inode == NULL);
+
+       while (block < last && block != EXT_MAX_BLOCK) {
+               num = last - block;
+               /* find extent for this block */
+               path = ext4_ext_find_extent(inode, block, path);
+               if (IS_ERR(path)) {
+                       err = PTR_ERR(path);
+                       path = NULL;
+                       break;
+               }
+
+               depth = ext_depth(inode);
+               BUG_ON(path[depth].p_hdr == NULL);
+               ex = path[depth].p_ext;
+               next = ext4_ext_next_allocated_block(path);
+
+               exists = 0;
+               if (!ex) {
+                       /* there is no extent yet, so try to allocate
+                        * all requested space */
+                       start = block;
+                       end = block + num;
+               } else if (le32_to_cpu(ex->ee_block) > block) {
+                       /* need to allocate space before found extent */
+                       start = block;
+                       end = le32_to_cpu(ex->ee_block);
+                       if (block + num < end)
+                               end = block + num;
+               } else if (block >=
+                            le32_to_cpu(ex->ee_block) + le16_to_cpu(ex->ee_len)) {
+                       /* need to allocate space after found extent */
+                       start = block;
+                       end = block + num;
+                       if (end >= next)
+                               end = next;
+               } else if (block >= le32_to_cpu(ex->ee_block)) {
+                       /*
+                        * some part of requested space is covered
+                        * by found extent
+                        */
+                       start = block;
+                       end = le32_to_cpu(ex->ee_block) + le16_to_cpu(ex->ee_len);
+                       if (block + num < end)
+                               end = block + num;
+                       exists = 1;
+               } else {
+                       BUG();
+               }
+               BUG_ON(end <= start);
+
+               if (!exists) {
+                       cbex.ec_block = start;
+                       cbex.ec_len = end - start;
+                       cbex.ec_start = 0;
+                       cbex.ec_type = EXT4_EXT_CACHE_GAP;
+               } else {
+                       cbex.ec_block = le32_to_cpu(ex->ee_block);
+                       cbex.ec_len = le16_to_cpu(ex->ee_len);
+                       cbex.ec_start = le32_to_cpu(ex->ee_start);
+                       cbex.ec_type = EXT4_EXT_CACHE_EXTENT;
+               }
+
+               BUG_ON(cbex.ec_len == 0);
+               err = func(inode, path, &cbex, cbdata);
+               ext4_ext_drop_refs(path);
+
+               if (err < 0)
+                       break;
+               if (err == EXT_REPEAT)
+                       continue;
+               else if (err == EXT_BREAK) {
+                       err = 0;
+                       break;
+               }
+
+               if (ext_depth(inode) != depth) {
+                       /* depth was changed. we have to realloc path */
+                       kfree(path);
+                       path = NULL;
+               }
+
+               block = cbex.ec_block + cbex.ec_len;
+       }
+
+       if (path) {
+               ext4_ext_drop_refs(path);
+               kfree(path);
+       }
+
+       return err;
+}
+
+static inline void
+ext4_ext_put_in_cache(struct inode *inode, __u32 block,
+                       __u32 len, __u32 start, int type)
+{
+       struct ext4_ext_cache *cex;
+       BUG_ON(len == 0);
+       cex = &EXT4_I(inode)->i_cached_extent;
+       cex->ec_type = type;
+       cex->ec_block = block;
+       cex->ec_len = len;
+       cex->ec_start = start;
+}
+
+/*
+ * this routine calculate boundaries of the gap requested block fits into
+ * and cache this gap
+ */
+static inline void
+ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
+                               unsigned long block)
+{
+       int depth = ext_depth(inode);
+       unsigned long lblock, len;
+       struct ext4_extent *ex;
+
+       ex = path[depth].p_ext;
+       if (ex == NULL) {
+               /* there is no extent yet, so gap is [0;-] */
+               lblock = 0;
+               len = EXT_MAX_BLOCK;
+               ext_debug("cache gap(whole file):");
+       } else if (block < le32_to_cpu(ex->ee_block)) {
+               lblock = block;
+               len = le32_to_cpu(ex->ee_block) - block;
+               ext_debug("cache gap(before): %lu [%lu:%lu]",
+                               (unsigned long) block,
+                               (unsigned long) le32_to_cpu(ex->ee_block),
+                               (unsigned long) le16_to_cpu(ex->ee_len));
+       } else if (block >= le32_to_cpu(ex->ee_block)
+                           + le16_to_cpu(ex->ee_len)) {
+               lblock = le32_to_cpu(ex->ee_block)
+                        + le16_to_cpu(ex->ee_len);
+               len = ext4_ext_next_allocated_block(path);
+               ext_debug("cache gap(after): [%lu:%lu] %lu",
+                               (unsigned long) le32_to_cpu(ex->ee_block),
+                               (unsigned long) le16_to_cpu(ex->ee_len),
+                               (unsigned long) block);
+               BUG_ON(len == lblock);
+               len = len - lblock;
+       } else {
+               lblock = len = 0;
+               BUG();
+       }
+
+       ext_debug(" -> %lu:%lu\n", (unsigned long) lblock, len);
+       ext4_ext_put_in_cache(inode, lblock, len, 0, EXT4_EXT_CACHE_GAP);
+}
+
+static inline int
+ext4_ext_in_cache(struct inode *inode, unsigned long block,
+                       struct ext4_extent *ex)
+{
+       struct ext4_ext_cache *cex;
+
+       cex = &EXT4_I(inode)->i_cached_extent;
+
+       /* has cache valid data? */
+       if (cex->ec_type == EXT4_EXT_CACHE_NO)
+               return EXT4_EXT_CACHE_NO;
+
+       BUG_ON(cex->ec_type != EXT4_EXT_CACHE_GAP &&
+                       cex->ec_type != EXT4_EXT_CACHE_EXTENT);
+       if (block >= cex->ec_block && block < cex->ec_block + cex->ec_len) {
+               ex->ee_block = cpu_to_le32(cex->ec_block);
+               ex->ee_start = cpu_to_le32(cex->ec_start);
+               ex->ee_len = cpu_to_le16(cex->ec_len);
+               ext_debug("%lu cached by %lu:%lu:%lu\n",
+                               (unsigned long) block,
+                               (unsigned long) cex->ec_block,
+                               (unsigned long) cex->ec_len,
+                               (unsigned long) cex->ec_start);
+               return cex->ec_type;
+       }
+
+       /* not in cache */
+       return EXT4_EXT_CACHE_NO;
+}
+
+/*
+ * routine removes index from the index block
+ * it's used in truncate case only. thus all requests are for
+ * last index in the block only
+ */
+int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
+                       struct ext4_ext_path *path)
+{
+       struct buffer_head *bh;
+       int err;
+       unsigned long leaf;
+
+       /* free index block */
+       path--;
+       leaf = le32_to_cpu(path->p_idx->ei_leaf);
+       BUG_ON(path->p_hdr->eh_entries == 0);
+       if ((err = ext4_ext_get_access(handle, inode, path)))
+               return err;
+       path->p_hdr->eh_entries = cpu_to_le16(le16_to_cpu(path->p_hdr->eh_entries)-1);
+       if ((err = ext4_ext_dirty(handle, inode, path)))
+               return err;
+       ext_debug("index is empty, remove it, free block %lu\n", leaf);
+       bh = sb_find_get_block(inode->i_sb, leaf);
+       ext4_forget(handle, 1, inode, bh, leaf);
+       ext4_free_blocks(handle, inode, leaf, 1);
+       return err;
+}
+
+/*
+ * This routine returns max. credits extent tree can consume.
+ * It should be OK for low-performance paths like ->writepage()
+ * To allow many writing process to fit a single transaction,
+ * caller should calculate credits under truncate_mutex and
+ * pass actual path.
+ */
+int inline ext4_ext_calc_credits_for_insert(struct inode *inode,
+                                               struct ext4_ext_path *path)
+{
+       int depth, needed;
+
+       if (path) {
+               /* probably there is space in leaf? */
+               depth = ext_depth(inode);
+               if (le16_to_cpu(path[depth].p_hdr->eh_entries)
+                               < le16_to_cpu(path[depth].p_hdr->eh_max))
+                       return 1;
+       }
+
+       /*
+        * given 32bit logical block (4294967296 blocks), max. tree
+        * can be 4 levels in depth -- 4 * 340^4 == 53453440000.
+        * let's also add one more level for imbalance.
+        */
+       depth = 5;
+
+       /* allocation of new data block(s) */
+       needed = 2;
+
+       /*
+        * tree can be full, so it'd need to grow in depth:
+        * allocation + old root + new root
+        */
+       needed += 2 + 1 + 1;
+
+       /*
+        * Index split can happen, we'd need:
+        *    allocate intermediate indexes (bitmap + group)
+        *  + change two blocks at each level, but root (already included)
+        */
+       needed = (depth * 2) + (depth * 2);
+
+       /* any allocation modifies superblock */
+       needed += 1;
+
+       return needed;
+}
+
+static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
+                               struct ext4_extent *ex,
+                               unsigned long from, unsigned long to)
+{
+       struct buffer_head *bh;
+       int i;
+
+#ifdef EXTENTS_STATS
+       {
+               struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
+               unsigned short ee_len =  le16_to_cpu(ex->ee_len);
+               spin_lock(&sbi->s_ext_stats_lock);
+               sbi->s_ext_blocks += ee_len;
+               sbi->s_ext_extents++;
+               if (ee_len < sbi->s_ext_min)
+                       sbi->s_ext_min = ee_len;
+               if (ee_len > sbi->s_ext_max)
+                       sbi->s_ext_max = ee_len;
+               if (ext_depth(inode) > sbi->s_depth_max)
+                       sbi->s_depth_max = ext_depth(inode);
+               spin_unlock(&sbi->s_ext_stats_lock);
+       }
+#endif
+       if (from >= le32_to_cpu(ex->ee_block)
+           && to == le32_to_cpu(ex->ee_block) + le16_to_cpu(ex->ee_len) - 1) {
+               /* tail removal */
+               unsigned long num, start;
+               num = le32_to_cpu(ex->ee_block) + le16_to_cpu(ex->ee_len) - from;
+               start = le32_to_cpu(ex->ee_start) + le16_to_cpu(ex->ee_len) - num;
+               ext_debug("free last %lu blocks starting %lu\n", num, start);
+               for (i = 0; i < num; i++) {
+                       bh = sb_find_get_block(inode->i_sb, start + i);
+                       ext4_forget(handle, 0, inode, bh, start + i);
+               }
+               ext4_free_blocks(handle, inode, start, num);
+       } else if (from == le32_to_cpu(ex->ee_block)
+                  && to <= le32_to_cpu(ex->ee_block) + le16_to_cpu(ex->ee_len) - 1) {
+               printk("strange request: removal %lu-%lu from %u:%u\n",
+                      from, to, le32_to_cpu(ex->ee_block), le16_to_cpu(ex->ee_len));
+       } else {
+               printk("strange request: removal(2) %lu-%lu from %u:%u\n",
+                      from, to, le32_to_cpu(ex->ee_block), le16_to_cpu(ex->ee_len));
+       }
+       return 0;
+}
+
+static int
+ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
+               struct ext4_ext_path *path, unsigned long start)
+{
+       int err = 0, correct_index = 0;
+       int depth = ext_depth(inode), credits;
+       struct ext4_extent_header *eh;
+       unsigned a, b, block, num;
+       unsigned long ex_ee_block;
+       unsigned short ex_ee_len;
+       struct ext4_extent *ex;
+
+       ext_debug("truncate since %lu in leaf\n", start);
+       if (!path[depth].p_hdr)
+               path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
+       eh = path[depth].p_hdr;
+       BUG_ON(eh == NULL);
+       BUG_ON(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max));
+       BUG_ON(eh->eh_magic != EXT4_EXT_MAGIC);
+
+       /* find where to start removing */
+       ex = EXT_LAST_EXTENT(eh);
+
+       ex_ee_block = le32_to_cpu(ex->ee_block);
+       ex_ee_len = le16_to_cpu(ex->ee_len);
+
+       while (ex >= EXT_FIRST_EXTENT(eh) &&
+                       ex_ee_block + ex_ee_len > start) {
+               ext_debug("remove ext %lu:%u\n", ex_ee_block, ex_ee_len);
+               path[depth].p_ext = ex;
+
+               a = ex_ee_block > start ? ex_ee_block : start;
+               b = ex_ee_block + ex_ee_len - 1 < EXT_MAX_BLOCK ?
+                       ex_ee_block + ex_ee_len - 1 : EXT_MAX_BLOCK;
+
+               ext_debug("  border %u:%u\n", a, b);
+
+               if (a != ex_ee_block && b != ex_ee_block + ex_ee_len - 1) {
+                       block = 0;
+                       num = 0;
+                       BUG();
+               } else if (a != ex_ee_block) {
+                       /* remove tail of the extent */
+                       block = ex_ee_block;
+                       num = a - block;
+               } else if (b != ex_ee_block + ex_ee_len - 1) {
+                       /* remove head of the extent */
+                       block = a;
+                       num = b - a;
+                       /* there is no "make a hole" API yet */
+                       BUG();
+               } else {
+                       /* remove whole extent: excellent! */
+                       block = ex_ee_block;
+                       num = 0;
+                       BUG_ON(a != ex_ee_block);
+                       BUG_ON(b != ex_ee_block + ex_ee_len - 1);
+               }
+
+               /* at present, extent can't cross block group */
+               /* leaf + bitmap + group desc + sb + inode */
+               credits = 5;
+               if (ex == EXT_FIRST_EXTENT(eh)) {
+                       correct_index = 1;
+                       credits += (ext_depth(inode)) + 1;
+               }
+#ifdef CONFIG_QUOTA
+               credits += 2 * EXT4_QUOTA_TRANS_BLOCKS(inode->i_sb);
+#endif
+
+               handle = ext4_ext_journal_restart(handle, credits);
+               if (IS_ERR(handle)) {
+                       err = PTR_ERR(handle);
+                       goto out;
+               }
+
+               err = ext4_ext_get_access(handle, inode, path + depth);
+               if (err)
+                       goto out;
+
+               err = ext4_remove_blocks(handle, inode, ex, a, b);
+               if (err)
+                       goto out;
+
+               if (num == 0) {
+                       /* this extent is removed entirely mark slot unused */
+                       ex->ee_start = 0;
+                       eh->eh_entries = cpu_to_le16(le16_to_cpu(eh->eh_entries)-1);
+               }
+
+               ex->ee_block = cpu_to_le32(block);
+               ex->ee_len = cpu_to_le16(num);
+
+               err = ext4_ext_dirty(handle, inode, path + depth);
+               if (err)
+                       goto out;
+
+               ext_debug("new extent: %u:%u:%u\n", block, num,
+                               le32_to_cpu(ex->ee_start));
+               ex--;
+               ex_ee_block = le32_to_cpu(ex->ee_block);
+               ex_ee_len = le16_to_cpu(ex->ee_len);
+       }
+
+       if (correct_index && eh->eh_entries)
+               err = ext4_ext_correct_indexes(handle, inode, path);
+
+       /* if this leaf is free, then we should
+        * remove it from index block above */
+       if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
+               err = ext4_ext_rm_idx(handle, inode, path + depth);
+
+out:
+       return err;
+}
+
+/*
+ * returns 1 if current index have to be freed (even partial)
+ */
+static int inline
+ext4_ext_more_to_rm(struct ext4_ext_path *path)
+{
+       BUG_ON(path->p_idx == NULL);
+
+       if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
+               return 0;
+
+       /*
+        * if truncate on deeper level happened it it wasn't partial
+        * so we have to consider current index for truncation
+        */
+       if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
+               return 0;
+       return 1;
+}
+
+int ext4_ext_remove_space(struct inode *inode, unsigned long start)
+{
+       struct super_block *sb = inode->i_sb;
+       int depth = ext_depth(inode);
+       struct ext4_ext_path *path;
+       handle_t *handle;
+       int i = 0, err = 0;
+
+       ext_debug("truncate since %lu\n", start);
+
+       /* probably first extent we're gonna free will be last in block */
+       handle = ext4_journal_start(inode, depth + 1);
+       if (IS_ERR(handle))
+               return PTR_ERR(handle);
+
+       ext4_ext_invalidate_cache(inode);
+
+       /*
+        * we start scanning from right side freeing all the blocks
+        * after i_size and walking into the deep
+        */
+       path = kmalloc(sizeof(struct ext4_ext_path) * (depth + 1), GFP_KERNEL);
+       if (path == NULL) {
+               ext4_journal_stop(handle);
+               return -ENOMEM;
+       }
+       memset(path, 0, sizeof(struct ext4_ext_path) * (depth + 1));
+       path[0].p_hdr = ext_inode_hdr(inode);
+       if (ext4_ext_check_header(__FUNCTION__, inode, path[0].p_hdr)) {
+               err = -EIO;
+               goto out;
+       }
+       path[0].p_depth = depth;
+
+       while (i >= 0 && err == 0) {
+               if (i == depth) {
+                       /* this is leaf block */
+                       err = ext4_ext_rm_leaf(handle, inode, path, start);
+                       /* root level have p_bh == NULL, brelse() eats this */
+                       brelse(path[i].p_bh);
+                       path[i].p_bh = NULL;
+                       i--;
+                       continue;
+               }
+
+               /* this is index block */
+               if (!path[i].p_hdr) {
+                       ext_debug("initialize header\n");
+                       path[i].p_hdr = ext_block_hdr(path[i].p_bh);
+                       if (ext4_ext_check_header(__FUNCTION__, inode,
+                                                       path[i].p_hdr)) {
+                               err = -EIO;
+                               goto out;
+                       }
+               }
+
+               BUG_ON(le16_to_cpu(path[i].p_hdr->eh_entries)
+                          > le16_to_cpu(path[i].p_hdr->eh_max));
+               BUG_ON(path[i].p_hdr->eh_magic != EXT4_EXT_MAGIC);
+
+               if (!path[i].p_idx) {
+                       /* this level hasn't touched yet */
+                       path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
+                       path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
+                       ext_debug("init index ptr: hdr 0x%p, num %d\n",
+                                 path[i].p_hdr,
+                                 le16_to_cpu(path[i].p_hdr->eh_entries));
+               } else {
+                       /* we've already was here, see at next index */
+                       path[i].p_idx--;
+               }
+
+               ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
+                               i, EXT_FIRST_INDEX(path[i].p_hdr),
+                               path[i].p_idx);
+               if (ext4_ext_more_to_rm(path + i)) {
+                       /* go to the next level */
+                       ext_debug("move to level %d (block %d)\n",
+                                 i + 1, le32_to_cpu(path[i].p_idx->ei_leaf));
+                       memset(path + i + 1, 0, sizeof(*path));
+                       path[i+1].p_bh =
+                               sb_bread(sb, le32_to_cpu(path[i].p_idx->ei_leaf));
+                       if (!path[i+1].p_bh) {
+                               /* should we reset i_size? */
+                               err = -EIO;
+                               break;
+                       }
+
+                       /* put actual number of indexes to know is this
+                        * number got changed at the next iteration */
+                       path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
+                       i++;
+               } else {
+                       /* we finish processing this index, go up */
+                       if (path[i].p_hdr->eh_entries == 0 && i > 0) {
+                               /* index is empty, remove it
+                                * handle must be already prepared by the
+                                * truncatei_leaf() */
+                               err = ext4_ext_rm_idx(handle, inode, path + i);
+                       }
+                       /* root level have p_bh == NULL, brelse() eats this */
+                       brelse(path[i].p_bh);
+                       path[i].p_bh = NULL;
+                       i--;
+                       ext_debug("return to level %d\n", i);
+               }
+       }
+
+       /* TODO: flexible tree reduction should be here */
+       if (path->p_hdr->eh_entries == 0) {
+               /*
+                * truncate to zero freed all the tree
+                * so, we need to correct eh_depth
+                */
+               err = ext4_ext_get_access(handle, inode, path);
+               if (err == 0) {
+                       ext_inode_hdr(inode)->eh_depth = 0;
+                       ext_inode_hdr(inode)->eh_max =
+                               cpu_to_le16(ext4_ext_space_root(inode));
+                       err = ext4_ext_dirty(handle, inode, path);
+               }
+       }
+out:
+       ext4_ext_tree_changed(inode);
+       ext4_ext_drop_refs(path);
+       kfree(path);
+       ext4_journal_stop(handle);
+
+       return err;
+}
+
+/*
+ * called at mount time
+ */
+void ext4_ext_init(struct super_block *sb)
+{
+       /*
+        * possible initialization would be here
+        */
+
+       if (test_opt(sb, EXTENTS)) {
+               printk("EXT4-fs: file extents enabled");
+#ifdef AGRESSIVE_TEST
+               printk(", agressive tests");
+#endif
+#ifdef CHECK_BINSEARCH
+               printk(", check binsearch");
+#endif
+#ifdef EXTENTS_STATS
+               printk(", stats");
+#endif
+               printk("\n");
+#ifdef EXTENTS_STATS
+               spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
+               EXT4_SB(sb)->s_ext_min = 1 << 30;
+               EXT4_SB(sb)->s_ext_max = 0;
+#endif
+       }
+}
+
+/*
+ * called at umount time
+ */
+void ext4_ext_release(struct super_block *sb)
+{
+       if (!test_opt(sb, EXTENTS))
+               return;
+
+#ifdef EXTENTS_STATS
+       if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
+               struct ext4_sb_info *sbi = EXT4_SB(sb);
+               printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
+                       sbi->s_ext_blocks, sbi->s_ext_extents,
+                       sbi->s_ext_blocks / sbi->s_ext_extents);
+               printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
+                       sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
+       }
+#endif
+}
+
+int ext4_ext_get_blocks(handle_t *handle, struct inode *inode, sector_t iblock,
+                       unsigned long max_blocks, struct buffer_head *bh_result,
+                       int create, int extend_disksize)
+{
+       struct ext4_ext_path *path = NULL;
+       struct ext4_extent newex, *ex;
+       int goal, newblock, err = 0, depth;
+       unsigned long allocated = 0;
+
+       __clear_bit(BH_New, &bh_result->b_state);
+       ext_debug("blocks %d/%lu requested for inode %u\n", (int) iblock,
+                       max_blocks, (unsigned) inode->i_ino);
+       mutex_lock(&EXT4_I(inode)->truncate_mutex);
+
+       /* check in cache */
+       if ((goal = ext4_ext_in_cache(inode, iblock, &newex))) {
+               if (goal == EXT4_EXT_CACHE_GAP) {
+                       if (!create) {
+                               /* block isn't allocated yet and
+                                * user don't want to allocate it */
+                               goto out2;
+                       }
+                       /* we should allocate requested block */
+               } else if (goal == EXT4_EXT_CACHE_EXTENT) {
+                       /* block is already allocated */
+                       newblock = iblock
+                                  - le32_to_cpu(newex.ee_block)
+                                  + le32_to_cpu(newex.ee_start);
+                       /* number of remain blocks in the extent */
+                       allocated = le16_to_cpu(newex.ee_len) -
+                                       (iblock - le32_to_cpu(newex.ee_block));
+                       goto out;
+               } else {
+                       BUG();
+               }
+       }
+
+       /* find extent for this block */
+       path = ext4_ext_find_extent(inode, iblock, NULL);
+       if (IS_ERR(path)) {
+               err = PTR_ERR(path);
+               path = NULL;
+               goto out2;
+       }
+
+       depth = ext_depth(inode);
+
+       /*
+        * consistent leaf must not be empty
+        * this situations is possible, though, _during_ tree modification
+        * this is why assert can't be put in ext4_ext_find_extent()
+        */
+       BUG_ON(path[depth].p_ext == NULL && depth != 0);
+
+       if ((ex = path[depth].p_ext)) {
+               unsigned long ee_block = le32_to_cpu(ex->ee_block);
+               unsigned long ee_start = le32_to_cpu(ex->ee_start);
+               unsigned short ee_len  = le16_to_cpu(ex->ee_len);
+               /* if found exent covers block, simple return it */
+               if (iblock >= ee_block && iblock < ee_block + ee_len) {
+                       newblock = iblock - ee_block + ee_start;
+                       /* number of remain blocks in the extent */
+                       allocated = ee_len - (iblock - ee_block);
+                       ext_debug("%d fit into %lu:%d -> %d\n", (int) iblock,
+                                       ee_block, ee_len, newblock);
+                       ext4_ext_put_in_cache(inode, ee_block, ee_len,
+                                               ee_start, EXT4_EXT_CACHE_EXTENT);
+                       goto out;
+               }
+       }
+
+       /*
+        * requested block isn't allocated yet
+        * we couldn't try to create block if create flag is zero
+        */
+       if (!create) {
+               /* put just found gap into cache to speedup subsequest reqs */
+               ext4_ext_put_gap_in_cache(inode, path, iblock);
+               goto out2;
+       }
+       /*
+         * Okay, we need to do block allocation.  Lazily initialize the block
+         * allocation info here if necessary
+        */
+       if (S_ISREG(inode->i_mode) && (!EXT4_I(inode)->i_block_alloc_info))
+               ext4_init_block_alloc_info(inode);
+
+       /* allocate new block */
+       goal = ext4_ext_find_goal(inode, path, iblock);
+       allocated = max_blocks;
+       newblock = ext4_new_blocks(handle, inode, goal, &allocated, &err);
+       if (!newblock)
+               goto out2;
+       ext_debug("allocate new block: goal %d, found %d/%lu\n",
+                       goal, newblock, allocated);
+
+       /* try to insert new extent into found leaf and return */
+       newex.ee_block = cpu_to_le32(iblock);
+       newex.ee_start = cpu_to_le32(newblock);
+       newex.ee_len = cpu_to_le16(allocated);
+       err = ext4_ext_insert_extent(handle, inode, path, &newex);
+       if (err)
+               goto out2;
+
+       if (extend_disksize && inode->i_size > EXT4_I(inode)->i_disksize)
+               EXT4_I(inode)->i_disksize = inode->i_size;
+
+       /* previous routine could use block we allocated */
+       newblock = le32_to_cpu(newex.ee_start);
+       __set_bit(BH_New, &bh_result->b_state);
+
+       ext4_ext_put_in_cache(inode, iblock, allocated, newblock,
+                               EXT4_EXT_CACHE_EXTENT);
+out:
+       if (allocated > max_blocks)
+               allocated = max_blocks;
+       ext4_ext_show_leaf(inode, path);
+       __set_bit(BH_Mapped, &bh_result->b_state);
+       bh_result->b_bdev = inode->i_sb->s_bdev;
+       bh_result->b_blocknr = newblock;
+out2:
+       if (path) {
+               ext4_ext_drop_refs(path);
+               kfree(path);
+       }
+       mutex_unlock(&EXT4_I(inode)->truncate_mutex);
+
+       return err ? err : allocated;
+}
+
+void ext4_ext_truncate(struct inode * inode, struct page *page)
+{
+       struct address_space *mapping = inode->i_mapping;
+       struct super_block *sb = inode->i_sb;
+       unsigned long last_block;
+       handle_t *handle;
+       int err = 0;
+
+       /*
+        * probably first extent we're gonna free will be last in block
+        */
+       err = ext4_writepage_trans_blocks(inode) + 3;
+       handle = ext4_journal_start(inode, err);
+       if (IS_ERR(handle)) {
+               if (page) {
+                       clear_highpage(page);
+                       flush_dcache_page(page);
+                       unlock_page(page);
+                       page_cache_release(page);
+               }
+               return;
+       }
+
+       if (page)
+               ext4_block_truncate_page(handle, page, mapping, inode->i_size);
+
+       mutex_lock(&EXT4_I(inode)->truncate_mutex);
+       ext4_ext_invalidate_cache(inode);
+
+       /*
+        * TODO: optimization is possible here
+        * probably we need not scaning at all,
+        * because page truncation is enough
+        */
+       if (ext4_orphan_add(handle, inode))
+               goto out_stop;
+
+       /* we have to know where to truncate from in crash case */
+       EXT4_I(inode)->i_disksize = inode->i_size;
+       ext4_mark_inode_dirty(handle, inode);
+
+       last_block = (inode->i_size + sb->s_blocksize - 1)
+                       >> EXT4_BLOCK_SIZE_BITS(sb);
+       err = ext4_ext_remove_space(inode, last_block);
+
+       /* In a multi-transaction truncate, we only make the final
+        * transaction synchronous */
+       if (IS_SYNC(inode))
+               handle->h_sync = 1;
+
+out_stop:
+       /*
+        * If this was a simple ftruncate(), and the file will remain alive
+        * then we need to clear up the orphan record which we created above.
+        * However, if this was a real unlink then we were called by
+        * ext4_delete_inode(), and we allow that function to clean up the
+        * orphan info for us.
+        */
+       if (inode->i_nlink)
+               ext4_orphan_del(handle, inode);
+
+       mutex_unlock(&EXT4_I(inode)->truncate_mutex);
+       ext4_journal_stop(handle);
+}
+
+/*
+ * this routine calculate max number of blocks we could modify
+ * in order to allocate new block for an inode
+ */
+int ext4_ext_writepage_trans_blocks(struct inode *inode, int num)
+{
+       int needed;
+
+       needed = ext4_ext_calc_credits_for_insert(inode, NULL);
+
+       /* caller want to allocate num blocks, but note it includes sb */
+       needed = needed * num - (num - 1);
+
+#ifdef CONFIG_QUOTA
+       needed += 2 * EXT4_QUOTA_TRANS_BLOCKS(inode->i_sb);
+#endif
+
+       return needed;
+}
+
+EXPORT_SYMBOL(ext4_mark_inode_dirty);
+EXPORT_SYMBOL(ext4_ext_invalidate_cache);
+EXPORT_SYMBOL(ext4_ext_insert_extent);
+EXPORT_SYMBOL(ext4_ext_walk_space);
+EXPORT_SYMBOL(ext4_ext_find_goal);
+EXPORT_SYMBOL(ext4_ext_calc_credits_for_insert);
+
index 34d39ae966f71102fc181cfb0800fdad29ce6709..e17a6c918d725ab117ec59ad5e1da315881d08e4 100644 (file)
@@ -615,6 +615,17 @@ got:
                ext4_std_error(sb, err);
                goto fail_free_drop;
        }
+       if (test_opt(sb, EXTENTS)) {
+               EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL;
+               ext4_ext_tree_init(handle, inode);
+               if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
+                       err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh);
+                       if (err) goto fail;
+                       EXT4_SET_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS);
+                       BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "call ext4_journal_dirty_metadata");
+                       err = ext4_journal_dirty_metadata(handle, EXT4_SB(sb)->s_sbh);
+               }
+       }
 
        ext4_debug("allocating inode %lu\n", inode->i_ino);
        goto really_out;
index 0d96c7d3bb5b9ce9a691995b492f2ff2701ab253..2b81b1324a6f4972b9833cf7abc8c9b152e9205e 100644 (file)
@@ -40,8 +40,6 @@
 #include "xattr.h"
 #include "acl.h"
 
-static int ext4_writepage_trans_blocks(struct inode *inode);
-
 /*
  * Test whether an inode is a fast symlink.
  */
@@ -804,6 +802,7 @@ int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
        ext4_fsblk_t first_block = 0;
 
 
+       J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
        J_ASSERT(handle != NULL || create == 0);
        depth = ext4_block_to_path(inode,iblock,offsets,&blocks_to_boundary);
 
@@ -984,7 +983,7 @@ static int ext4_get_block(struct inode *inode, sector_t iblock,
 
 get_block:
        if (ret == 0) {
-               ret = ext4_get_blocks_handle(handle, inode, iblock,
+               ret = ext4_get_blocks_wrap(handle, inode, iblock,
                                        max_blocks, bh_result, create, 0);
                if (ret > 0) {
                        bh_result->b_size = (ret << inode->i_blkbits);
@@ -1008,7 +1007,7 @@ struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
        dummy.b_state = 0;
        dummy.b_blocknr = -1000;
        buffer_trace_init(&dummy.b_history);
-       err = ext4_get_blocks_handle(handle, inode, block, 1,
+       err = ext4_get_blocks_wrap(handle, inode, block, 1,
                                        &dummy, create, 1);
        /*
         * ext4_get_blocks_handle() returns number of blocks
@@ -1759,7 +1758,7 @@ void ext4_set_aops(struct inode *inode)
  * This required during truncate. We need to physically zero the tail end
  * of that block so it doesn't yield old data if the file is later grown.
  */
-static int ext4_block_truncate_page(handle_t *handle, struct page *page,
+int ext4_block_truncate_page(handle_t *handle, struct page *page,
                struct address_space *mapping, loff_t from)
 {
        ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
@@ -2263,6 +2262,9 @@ void ext4_truncate(struct inode *inode)
                        return;
        }
 
+       if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
+               return ext4_ext_truncate(inode, page);
+
        handle = start_transaction(inode);
        if (IS_ERR(handle)) {
                if (page) {
@@ -3003,12 +3005,15 @@ err_out:
  * block and work out the exact number of indirects which are touched.  Pah.
  */
 
-static int ext4_writepage_trans_blocks(struct inode *inode)
+int ext4_writepage_trans_blocks(struct inode *inode)
 {
        int bpp = ext4_journal_blocks_per_page(inode);
        int indirects = (EXT4_NDIR_BLOCKS % bpp) ? 5 : 3;
        int ret;
 
+       if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
+               return ext4_ext_writepage_trans_blocks(inode, bpp);
+
        if (ext4_should_journal_data(inode))
                ret = 3 * (bpp + indirects) + 2;
        else
index a63dce2117b8e9d9f08d4d8e38dbed9b45eaa358..22a737c306c7846693e4b032f2ca71091ec1113c 100644 (file)
@@ -248,7 +248,6 @@ flags_err:
                return err;
        }
 
-
        default:
                return -ENOTTY;
        }
index f131bb69b62e3be26abf0601fe2b5427305bcebe..69f8752505003f263b7bf71fbf470ae697a9321d 100644 (file)
@@ -390,6 +390,7 @@ static void ext4_put_super (struct super_block * sb)
        struct ext4_super_block *es = sbi->s_es;
        int i;
 
+       ext4_ext_release(sb);
        ext4_xattr_put_super(sb);
        jbd2_journal_destroy(sbi->s_journal);
        if (!(sb->s_flags & MS_RDONLY)) {
@@ -454,6 +455,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
 #endif
        ei->i_block_alloc_info = NULL;
        ei->vfs_inode.i_version = 1;
+       memset(&ei->i_cached_extent, 0, sizeof(struct ext4_ext_cache));
        return &ei->vfs_inode;
 }
 
@@ -677,7 +679,7 @@ enum {
        Opt_usrjquota, Opt_grpjquota, Opt_offusrjquota, Opt_offgrpjquota,
        Opt_jqfmt_vfsold, Opt_jqfmt_vfsv0, Opt_quota, Opt_noquota,
        Opt_ignore, Opt_barrier, Opt_err, Opt_resize, Opt_usrquota,
-       Opt_grpquota
+       Opt_grpquota, Opt_extents,
 };
 
 static match_table_t tokens = {
@@ -727,6 +729,7 @@ static match_table_t tokens = {
        {Opt_quota, "quota"},
        {Opt_usrquota, "usrquota"},
        {Opt_barrier, "barrier=%u"},
+       {Opt_extents, "extents"},
        {Opt_err, NULL},
        {Opt_resize, "resize"},
 };
@@ -1059,6 +1062,9 @@ clear_qf_name:
                case Opt_bh:
                        clear_opt(sbi->s_mount_opt, NOBH);
                        break;
+               case Opt_extents:
+                       set_opt (sbi->s_mount_opt, EXTENTS);
+                       break;
                default:
                        printk (KERN_ERR
                                "EXT4-fs: Unrecognized mount option \"%s\" "
@@ -1787,6 +1793,8 @@ static int ext4_fill_super (struct super_block *sb, void *data, int silent)
                test_opt(sb,DATA_FLAGS) == EXT4_MOUNT_ORDERED_DATA ? "ordered":
                "writeback");
 
+       ext4_ext_init(sb);
+
        lock_kernel();
        return 0;
 
index f582cd762cafd880b30ca146fdd6707ccf7310ab..b61181aadcbbcbcd332d6c41eaaf78c8ad7f35e2 100644 (file)
@@ -178,8 +178,9 @@ struct ext4_group_desc
 #define EXT4_DIRSYNC_FL                        0x00010000 /* dirsync behaviour (directories only) */
 #define EXT4_TOPDIR_FL                 0x00020000 /* Top of directory hierarchies*/
 #define EXT4_RESERVED_FL               0x80000000 /* reserved for ext4 lib */
+#define EXT4_EXTENTS_FL                        0x00080000 /* Inode uses extents */
 
-#define EXT4_FL_USER_VISIBLE           0x0003DFFF /* User visible flags */
+#define EXT4_FL_USER_VISIBLE           0x000BDFFF /* User visible flags */
 #define EXT4_FL_USER_MODIFIABLE                0x000380FF /* User modifiable flags */
 
 /*
@@ -384,6 +385,7 @@ struct ext4_inode {
 #define EXT4_MOUNT_QUOTA               0x80000 /* Some quota option set */
 #define EXT4_MOUNT_USRQUOTA            0x100000 /* "old" user quota */
 #define EXT4_MOUNT_GRPQUOTA            0x200000 /* "old" group quota */
+#define EXT4_MOUNT_EXTENTS             0x400000 /* Extents support */
 
 /* Compatibility, for having both ext2_fs.h and ext4_fs.h included at once */
 #ifndef _LINUX_EXT2_FS_H
@@ -582,11 +584,13 @@ static inline int ext4_valid_inum(struct super_block *sb, unsigned long ino)
 #define EXT4_FEATURE_INCOMPAT_RECOVER          0x0004 /* Needs recovery */
 #define EXT4_FEATURE_INCOMPAT_JOURNAL_DEV      0x0008 /* Journal device */
 #define EXT4_FEATURE_INCOMPAT_META_BG          0x0010
+#define EXT4_FEATURE_INCOMPAT_EXTENTS          0x0040 /* extents support */
 
 #define EXT4_FEATURE_COMPAT_SUPP       EXT2_FEATURE_COMPAT_EXT_ATTR
 #define EXT4_FEATURE_INCOMPAT_SUPP     (EXT4_FEATURE_INCOMPAT_FILETYPE| \
                                         EXT4_FEATURE_INCOMPAT_RECOVER| \
-                                        EXT4_FEATURE_INCOMPAT_META_BG)
+                                        EXT4_FEATURE_INCOMPAT_META_BG| \
+                                        EXT4_FEATURE_INCOMPAT_EXTENTS)
 #define EXT4_FEATURE_RO_COMPAT_SUPP    (EXT4_FEATURE_RO_COMPAT_SPARSE_SUPER| \
                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE| \
                                         EXT4_FEATURE_RO_COMPAT_BTREE_DIR)
@@ -825,6 +829,9 @@ extern int ext4_get_inode_loc(struct inode *, struct ext4_iloc *);
 extern void ext4_truncate (struct inode *);
 extern void ext4_set_inode_flags(struct inode *);
 extern void ext4_set_aops(struct inode *inode);
+extern int ext4_writepage_trans_blocks(struct inode *);
+extern int ext4_block_truncate_page(handle_t *handle, struct page *page,
+               struct address_space *mapping, loff_t from);
 
 /* ioctl.c */
 extern int ext4_ioctl (struct inode *, struct file *, unsigned int,
@@ -879,6 +886,28 @@ extern struct inode_operations ext4_special_inode_operations;
 extern struct inode_operations ext4_symlink_inode_operations;
 extern struct inode_operations ext4_fast_symlink_inode_operations;
 
+/* extents.c */
+extern int ext4_ext_tree_init(handle_t *handle, struct inode *);
+extern int ext4_ext_writepage_trans_blocks(struct inode *, int);
+extern int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
+                       ext4_fsblk_t iblock,
+                       unsigned long max_blocks, struct buffer_head *bh_result,
+                       int create, int extend_disksize);
+extern void ext4_ext_truncate(struct inode *, struct page *);
+extern void ext4_ext_init(struct super_block *);
+extern void ext4_ext_release(struct super_block *);
+static inline int
+ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
+                       unsigned long max_blocks, struct buffer_head *bh,
+                       int create, int extend_disksize)
+{
+       if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
+               return ext4_ext_get_blocks(handle, inode, block, max_blocks,
+                                       bh, create, extend_disksize);
+       return ext4_get_blocks_handle(handle, inode, block, max_blocks, bh,
+                                       create, extend_disksize);
+}
+
 
 #endif /* __KERNEL__ */
 
diff --git a/include/linux/ext4_fs_extents.h b/include/linux/ext4_fs_extents.h
new file mode 100644 (file)
index 0000000..8029879
--- /dev/null
@@ -0,0 +1,196 @@
+/*
+ * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
+ * Written by Alex Tomas <alex@clusterfs.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public Licens
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
+ */
+
+#ifndef _LINUX_EXT4_EXTENTS
+#define _LINUX_EXT4_EXTENTS
+
+#include <linux/ext4_fs.h>
+
+/*
+ * with AGRESSIVE_TEST defined capacity of index/leaf blocks
+ * become very little, so index split, in-depth growing and
+ * other hard changes happens much more often
+ * this is for debug purposes only
+ */
+#define AGRESSIVE_TEST_
+
+/*
+ * with EXTENTS_STATS defined number of blocks and extents
+ * are collected in truncate path. they'll be showed at
+ * umount time
+ */
+#define EXTENTS_STATS__
+
+/*
+ * if CHECK_BINSEARCH defined, then results of binary search
+ * will be checked by linear search
+ */
+#define CHECK_BINSEARCH__
+
+/*
+ * if EXT_DEBUG is defined you can use 'extdebug' mount option
+ * to get lots of info what's going on
+ */
+#define EXT_DEBUG__
+#ifdef EXT_DEBUG
+#define ext_debug(a...)                printk(a)
+#else
+#define ext_debug(a...)
+#endif
+
+/*
+ * if EXT_STATS is defined then stats numbers are collected
+ * these number will be displayed at umount time
+ */
+#define EXT_STATS_
+
+
+/*
+ * ext4_inode has i_block array (60 bytes total)
+ * first 12 bytes store ext4_extent_header
+ * the remain stores array of ext4_extent
+ */
+
+/*
+ * this is extent on-disk structure
+ * it's used at the bottom of the tree
+ */
+struct ext4_extent {
+       __le32  ee_block;       /* first logical block extent covers */
+       __le16  ee_len;         /* number of blocks covered by extent */
+       __le16  ee_start_hi;    /* high 16 bits of physical block */
+       __le32  ee_start;       /* low 32 bigs of physical block */
+};
+
+/*
+ * this is index on-disk structure
+ * it's used at all the levels, but the bottom
+ */
+struct ext4_extent_idx {
+       __le32  ei_block;       /* index covers logical blocks from 'block' */
+       __le32  ei_leaf;        /* pointer to the physical block of the next *
+                                * level. leaf or next index could bet here */
+       __le16  ei_leaf_hi;     /* high 16 bits of physical block */
+       __u16   ei_unused;
+};
+
+/*
+ * each block (leaves and indexes), even inode-stored has header
+ */
+struct ext4_extent_header {
+       __le16  eh_magic;       /* probably will support different formats */
+       __le16  eh_entries;     /* number of valid entries */
+       __le16  eh_max;         /* capacity of store in entries */
+       __le16  eh_depth;       /* has tree real underlaying blocks? */
+       __le32  eh_generation;  /* generation of the tree */
+};
+
+#define EXT4_EXT_MAGIC         cpu_to_le16(0xf30a)
+
+/*
+ * array of ext4_ext_path contains path to some extent
+ * creation/lookup routines use it for traversal/splitting/etc
+ * truncate uses it to simulate recursive walking
+ */
+struct ext4_ext_path {
+       __u32                           p_block;
+       __u16                           p_depth;
+       struct ext4_extent              *p_ext;
+       struct ext4_extent_idx          *p_idx;
+       struct ext4_extent_header       *p_hdr;
+       struct buffer_head              *p_bh;
+};
+
+/*
+ * structure for external API
+ */
+
+#define EXT4_EXT_CACHE_NO      0
+#define EXT4_EXT_CACHE_GAP     1
+#define EXT4_EXT_CACHE_EXTENT  2
+
+/*
+ * to be called by ext4_ext_walk_space()
+ * negative retcode - error
+ * positive retcode - signal for ext4_ext_walk_space(), see below
+ * callback must return valid extent (passed or newly created)
+ */
+typedef int (*ext_prepare_callback)(struct inode *, struct ext4_ext_path *,
+                                       struct ext4_ext_cache *,
+                                       void *);
+
+#define EXT_CONTINUE   0
+#define EXT_BREAK      1
+#define EXT_REPEAT     2
+
+
+#define EXT_MAX_BLOCK  0xffffffff
+
+
+#define EXT_FIRST_EXTENT(__hdr__) \
+       ((struct ext4_extent *) (((char *) (__hdr__)) +         \
+                                sizeof(struct ext4_extent_header)))
+#define EXT_FIRST_INDEX(__hdr__) \
+       ((struct ext4_extent_idx *) (((char *) (__hdr__)) +     \
+                                    sizeof(struct ext4_extent_header)))
+#define EXT_HAS_FREE_INDEX(__path__) \
+        (le16_to_cpu((__path__)->p_hdr->eh_entries) \
+                                    < le16_to_cpu((__path__)->p_hdr->eh_max))
+#define EXT_LAST_EXTENT(__hdr__) \
+       (EXT_FIRST_EXTENT((__hdr__)) + le16_to_cpu((__hdr__)->eh_entries) - 1)
+#define EXT_LAST_INDEX(__hdr__) \
+       (EXT_FIRST_INDEX((__hdr__)) + le16_to_cpu((__hdr__)->eh_entries) - 1)
+#define EXT_MAX_EXTENT(__hdr__) \
+       (EXT_FIRST_EXTENT((__hdr__)) + le16_to_cpu((__hdr__)->eh_max) - 1)
+#define EXT_MAX_INDEX(__hdr__) \
+       (EXT_FIRST_INDEX((__hdr__)) + le16_to_cpu((__hdr__)->eh_max) - 1)
+
+static inline struct ext4_extent_header *ext_inode_hdr(struct inode *inode)
+{
+       return (struct ext4_extent_header *) EXT4_I(inode)->i_data;
+}
+
+static inline struct ext4_extent_header *ext_block_hdr(struct buffer_head *bh)
+{
+       return (struct ext4_extent_header *) bh->b_data;
+}
+
+static inline unsigned short ext_depth(struct inode *inode)
+{
+       return le16_to_cpu(ext_inode_hdr(inode)->eh_depth);
+}
+
+static inline void ext4_ext_tree_changed(struct inode *inode)
+{
+       EXT4_I(inode)->i_ext_generation++;
+}
+
+static inline void
+ext4_ext_invalidate_cache(struct inode *inode)
+{
+       EXT4_I(inode)->i_cached_extent.ec_type = EXT4_EXT_CACHE_NO;
+}
+
+extern int ext4_extent_tree_init(handle_t *, struct inode *);
+extern int ext4_ext_calc_credits_for_insert(struct inode *, struct ext4_ext_path *);
+extern int ext4_ext_insert_extent(handle_t *, struct inode *, struct ext4_ext_path *, struct ext4_extent *);
+extern int ext4_ext_walk_space(struct inode *, unsigned long, unsigned long, ext_prepare_callback, void *);
+extern struct ext4_ext_path * ext4_ext_find_extent(struct inode *, int, struct ext4_ext_path *);
+
+#endif /* _LINUX_EXT4_EXTENTS */
+
index 18a6ce98537f5d033ac1a2abb6c2937ac0ca9148..40ce04a52b04c1da43cbf541f6f671ed066a2739 100644 (file)
@@ -64,6 +64,16 @@ struct ext4_block_alloc_info {
 #define rsv_start rsv_window._rsv_start
 #define rsv_end rsv_window._rsv_end
 
+/*
+ * storage for cached extent
+ */
+struct ext4_ext_cache {
+       __u32   ec_start;
+       __u32   ec_block;
+       __u32   ec_len; /* must be 32bit to return holes */
+       __u32   ec_type;
+};
+
 /*
  * third extended file system inode data in memory
  */
@@ -142,6 +152,9 @@ struct ext4_inode_info {
         */
        struct mutex truncate_mutex;
        struct inode vfs_inode;
+
+       unsigned long i_ext_generation;
+       struct ext4_ext_cache i_cached_extent;
 };
 
 #endif /* _LINUX_EXT4_FS_I */
index ce4856d7010064acac77f5d09fdefa138f7bcfc4..ce7a844d670392dfb9d878617fd3f45f7b9b9c26 100644 (file)
@@ -78,6 +78,16 @@ struct ext4_sb_info {
        char *s_qf_names[MAXQUOTAS];            /* Names of quota files with journalled quota */
        int s_jquota_fmt;                       /* Format of quota to use */
 #endif
+
+#ifdef EXTENTS_STATS
+       /* ext4 extents stats */
+       unsigned long s_ext_min;
+       unsigned long s_ext_max;
+       unsigned long s_depth_max;
+       spinlock_t s_ext_stats_lock;
+       unsigned long s_ext_blocks;
+       unsigned long s_ext_extents;
+#endif
 };
 
 #endif /* _LINUX_EXT4_FS_SB */
index 99d37557cbb4c891fe5aba83e3d05a838a0f8875..aa273f024ad0b1376601df54437ab7935632b87d 100644 (file)
  *
  * We may have to touch one inode, one bitmap buffer, up to three
  * indirection blocks, the group and superblock summaries, and the data
- * block to complete the transaction.  */
+ * block to complete the transaction.
+ *
+ * For extents-enabled fs we may have to allocate and modify upto
+ * 5 levels of tree + root which is stored in inode. */
 
-#define EXT4_SINGLEDATA_TRANS_BLOCKS   8U
+#define EXT4_SINGLEDATA_TRANS_BLOCKS(sb)                               \
+       (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)   \
+               || test_opt(sb, EXTENTS) ? 27U : 8U)
 
 /* Extended attribute operations touch at most two data buffers,
  * two bitmap buffers, and two group summaries, in addition to the inode
@@ -42,7 +47,7 @@
  * superblock only gets updated once, of course, so don't bother
  * counting that again for the quota updates. */
 
-#define EXT4_DATA_TRANS_BLOCKS(sb)     (EXT4_SINGLEDATA_TRANS_BLOCKS + \
+#define EXT4_DATA_TRANS_BLOCKS(sb)     (EXT4_SINGLEDATA_TRANS_BLOCKS(sb) + \
                                         EXT4_XATTR_TRANS_BLOCKS - 2 + \
                                         2*EXT4_QUOTA_TRANS_BLOCKS(sb))
 
@@ -78,9 +83,9 @@
 /* Amount of blocks needed for quota insert/delete - we do some block writes
  * but inode, sb and group updates are done only once */
 #define EXT4_QUOTA_INIT_BLOCKS(sb) (test_opt(sb, QUOTA) ? (DQUOT_INIT_ALLOC*\
-               (EXT4_SINGLEDATA_TRANS_BLOCKS-3)+3+DQUOT_INIT_REWRITE) : 0)
+               (EXT4_SINGLEDATA_TRANS_BLOCKS(sb)-3)+3+DQUOT_INIT_REWRITE) : 0)
 #define EXT4_QUOTA_DEL_BLOCKS(sb) (test_opt(sb, QUOTA) ? (DQUOT_DEL_ALLOC*\
-               (EXT4_SINGLEDATA_TRANS_BLOCKS-3)+3+DQUOT_DEL_REWRITE) : 0)
+               (EXT4_SINGLEDATA_TRANS_BLOCKS(sb)-3)+3+DQUOT_DEL_REWRITE) : 0)
 #else
 #define EXT4_QUOTA_TRANS_BLOCKS(sb) 0
 #define EXT4_QUOTA_INIT_BLOCKS(sb) 0