ext4: Add multi block allocator for ext4
authorAlex Tomas <alex@clusterfs.com>
Tue, 29 Jan 2008 05:19:52 +0000 (00:19 -0500)
committerTheodore Ts'o <tytso@mit.edu>
Tue, 29 Jan 2008 05:19:52 +0000 (00:19 -0500)
Signed-off-by: Alex Tomas <alex@clusterfs.com>
Signed-off-by: Andreas Dilger <adilger@clusterfs.com>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
13 files changed:
Documentation/filesystems/ext4.txt
Documentation/filesystems/proc.txt
fs/ext4/Makefile
fs/ext4/balloc.c
fs/ext4/extents.c
fs/ext4/inode.c
fs/ext4/mballoc.c [new file with mode: 0644]
fs/ext4/migrate.c
fs/ext4/super.c
fs/ext4/xattr.c
include/linux/ext4_fs.h
include/linux/ext4_fs_i.h
include/linux/ext4_fs_sb.h

index 4f329af..560f88d 100644 (file)
@@ -86,9 +86,11 @@ Alex is working on a new set of patches right now.
 When mounting an ext4 filesystem, the following option are accepted:
 (*) == default
 
-extents                        ext4 will use extents to address file data.  The
+extents                (*)     ext4 will use extents to address file data.  The
                        file system will no longer be mountable by ext3.
 
+noextents              ext4 will not use extents for newly created files
+
 journal_checksum       Enable checksumming of the journal transactions.
                        This will allow the recovery code in e2fsck and the
                        kernel to detect corruption in the kernel.  It is a
@@ -206,6 +208,12 @@ nobh                       (a) cache disk block mapping information
                        "nobh" option tries to avoid associating buffer
                        heads (supported only for "writeback" mode).
 
+mballoc                (*)     Use the multiple block allocator for block allocation
+nomballoc              disabled multiple block allocator for block allocation.
+stripe=n               Number of filesystem blocks that mballoc will try
+                       to use for allocation size and alignment. For RAID5/6
+                       systems this should be the number of data
+                       disks *  RAID chunk size in file system blocks.
 
 Data Mode
 ---------
index dec9945..4413a2d 100644 (file)
@@ -857,6 +857,45 @@ CPUs.
 The   "procs_blocked" line gives  the  number of  processes currently blocked,
 waiting for I/O to complete.
 
+1.9 Ext4 file system parameters
+------------------------------
+Ext4 file system have one directory per partition under /proc/fs/ext4/
+# ls /proc/fs/ext4/hdc/
+group_prealloc  max_to_scan  mb_groups  mb_history  min_to_scan  order2_req
+stats  stream_req
+
+mb_groups:
+This file gives the details of mutiblock allocator buddy cache of free blocks
+
+mb_history:
+Multiblock allocation history.
+
+stats:
+This file indicate whether the multiblock allocator should start collecting
+statistics. The statistics are shown during unmount
+
+group_prealloc:
+The multiblock allocator normalize the block allocation request to
+group_prealloc filesystem blocks if we don't have strip value set.
+The stripe value can be specified at mount time or during mke2fs.
+
+max_to_scan:
+How long multiblock allocator can look for a best extent (in found extents)
+
+min_to_scan:
+How long multiblock allocator  must look for a best extent
+
+order2_req:
+Multiblock allocator use  2^N search using buddies only for requests greater
+than or equal to order2_req. The request size is specfied in file system
+blocks. A value of 2 indicate only if the requests are greater than or equal
+to 4 blocks.
+
+stream_req:
+Files smaller than stream_req are served by the stream allocator, whose
+purpose is to pack requests as close each to other as possible to
+produce smooth I/O traffic. Avalue of 16 indicate that file smaller than 16
+filesystem block size will use group based preallocation.
 
 ------------------------------------------------------------------------------
 Summary
index d5fd80b..ac6fa8c 100644 (file)
@@ -6,7 +6,7 @@ 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 extents.o \
-                  ext4_jbd2.o migrate.o
+                  ext4_jbd2.o migrate.o mballoc.o
 
 ext4dev-$(CONFIG_EXT4DEV_FS_XATTR)     += xattr.o xattr_user.o xattr_trusted.o
 ext4dev-$(CONFIG_EXT4DEV_FS_POSIX_ACL) += acl.o
index 80a4616..ac75ea9 100644 (file)
@@ -577,6 +577,8 @@ void ext4_discard_reservation(struct inode *inode)
        struct ext4_reserve_window_node *rsv;
        spinlock_t *rsv_lock = &EXT4_SB(inode->i_sb)->s_rsv_window_lock;
 
+       ext4_mb_discard_inode_preallocations(inode);
+
        if (!block_i)
                return;
 
@@ -785,19 +787,29 @@ error_return:
  * @inode:             inode
  * @block:             start physical block to free
  * @count:             number of blocks to count
+ * @metadata:          Are these metadata blocks
  */
 void ext4_free_blocks(handle_t *handle, struct inode *inode,
-                       ext4_fsblk_t block, unsigned long count)
+                       ext4_fsblk_t block, unsigned long count,
+                       int metadata)
 {
        struct super_block * sb;
        unsigned long dquot_freed_blocks;
 
+       /* this isn't the right place to decide whether block is metadata
+        * inode.c/extents.c knows better, but for safety ... */
+       if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
+                       ext4_should_journal_data(inode))
+               metadata = 1;
+
        sb = inode->i_sb;
-       if (!sb) {
-               printk ("ext4_free_blocks: nonexistent device");
-               return;
-       }
-       ext4_free_blocks_sb(handle, sb, block, count, &dquot_freed_blocks);
+
+       if (!test_opt(sb, MBALLOC) || !EXT4_SB(sb)->s_group_info)
+               ext4_free_blocks_sb(handle, sb, block, count,
+                                               &dquot_freed_blocks);
+       else
+               ext4_mb_free_blocks(handle, inode, block, count,
+                                               metadata, &dquot_freed_blocks);
        if (dquot_freed_blocks)
                DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
        return;
@@ -1576,7 +1588,7 @@ int ext4_should_retry_alloc(struct super_block *sb, int *retries)
 }
 
 /**
- * ext4_new_blocks() -- core block(s) allocation function
+ * ext4_new_blocks_old() -- core block(s) allocation function
  * @handle:            handle to this transaction
  * @inode:             file inode
  * @goal:              given target block(filesystem wide)
@@ -1589,7 +1601,7 @@ int ext4_should_retry_alloc(struct super_block *sb, int *retries)
  * any specific goal block.
  *
  */
-ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
+ext4_fsblk_t ext4_new_blocks_old(handle_t *handle, struct inode *inode,
                        ext4_fsblk_t goal, unsigned long *count, int *errp)
 {
        struct buffer_head *bitmap_bh = NULL;
@@ -1849,13 +1861,46 @@ out:
 }
 
 ext4_fsblk_t ext4_new_block(handle_t *handle, struct inode *inode,
-                       ext4_fsblk_t goal, int *errp)
+               ext4_fsblk_t goal, int *errp)
+{
+       struct ext4_allocation_request ar;
+       ext4_fsblk_t ret;
+
+       if (!test_opt(inode->i_sb, MBALLOC)) {
+               unsigned long count = 1;
+               ret = ext4_new_blocks_old(handle, inode, goal, &count, errp);
+               return ret;
+       }
+
+       memset(&ar, 0, sizeof(ar));
+       ar.inode = inode;
+       ar.goal = goal;
+       ar.len = 1;
+       ret = ext4_mb_new_blocks(handle, &ar, errp);
+       return ret;
+}
+
+ext4_fsblk_t ext4_new_blocks(handle_t *handle, struct inode *inode,
+               ext4_fsblk_t goal, unsigned long *count, int *errp)
 {
-       unsigned long count = 1;
+       struct ext4_allocation_request ar;
+       ext4_fsblk_t ret;
 
-       return ext4_new_blocks(handle, inode, goal, &count, errp);
+       if (!test_opt(inode->i_sb, MBALLOC)) {
+               ret = ext4_new_blocks_old(handle, inode, goal, count, errp);
+               return ret;
+       }
+
+       memset(&ar, 0, sizeof(ar));
+       ar.inode = inode;
+       ar.goal = goal;
+       ar.len = *count;
+       ret = ext4_mb_new_blocks(handle, &ar, errp);
+       *count = ar.len;
+       return ret;
 }
 
+
 /**
  * ext4_count_free_blocks() -- count filesystem free blocks
  * @sb:                superblock
index f5cf2a9..0cffb59 100644 (file)
@@ -853,7 +853,7 @@ cleanup:
                for (i = 0; i < depth; i++) {
                        if (!ablocks[i])
                                continue;
-                       ext4_free_blocks(handle, inode, ablocks[i], 1);
+                       ext4_free_blocks(handle, inode, ablocks[i], 1, 1);
                }
        }
        kfree(ablocks);
@@ -1698,7 +1698,7 @@ static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
        ext_debug("index is empty, remove it, free block %llu\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);
+       ext4_free_blocks(handle, inode, leaf, 1, 1);
        return err;
 }
 
@@ -1759,8 +1759,10 @@ static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
 {
        struct buffer_head *bh;
        unsigned short ee_len =  ext4_ext_get_actual_len(ex);
-       int i;
+       int i, metadata = 0;
 
+       if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
+               metadata = 1;
 #ifdef EXTENTS_STATS
        {
                struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
@@ -1789,7 +1791,7 @@ static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
                        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);
+               ext4_free_blocks(handle, inode, start, num, metadata);
        } else if (from == le32_to_cpu(ex->ee_block)
                   && to <= le32_to_cpu(ex->ee_block) + ee_len - 1) {
                printk(KERN_INFO "strange request: removal %u-%u from %u:%u\n",
@@ -2287,6 +2289,7 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
        ext4_fsblk_t goal, newblock;
        int err = 0, depth, ret;
        unsigned long allocated = 0;
+       struct ext4_allocation_request ar;
 
        __clear_bit(BH_New, &bh_result->b_state);
        ext_debug("blocks %u/%lu requested for inode %u\n",
@@ -2397,8 +2400,15 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
        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);
+       /* find neighbour allocated blocks */
+       ar.lleft = iblock;
+       err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
+       if (err)
+               goto out2;
+       ar.lright = iblock;
+       err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright);
+       if (err)
+               goto out2;
 
        /*
         * See if request is beyond maximum number of blocks we can have in
@@ -2421,7 +2431,18 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
                allocated = le16_to_cpu(newex.ee_len);
        else
                allocated = max_blocks;
-       newblock = ext4_new_blocks(handle, inode, goal, &allocated, &err);
+
+       /* allocate new block */
+       ar.inode = inode;
+       ar.goal = ext4_ext_find_goal(inode, path, iblock);
+       ar.logical = iblock;
+       ar.len = allocated;
+       if (S_ISREG(inode->i_mode))
+               ar.flags = EXT4_MB_HINT_DATA;
+       else
+               /* disable in-core preallocation for non-regular files */
+               ar.flags = 0;
+       newblock = ext4_mb_new_blocks(handle, &ar, &err);
        if (!newblock)
                goto out2;
        ext_debug("allocate new block: goal %llu, found %llu/%lu\n",
@@ -2429,14 +2450,17 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
 
        /* try to insert new extent into found leaf and return */
        ext4_ext_store_pblock(&newex, newblock);
-       newex.ee_len = cpu_to_le16(allocated);
+       newex.ee_len = cpu_to_le16(ar.len);
        if (create == EXT4_CREATE_UNINITIALIZED_EXT)  /* Mark uninitialized */
                ext4_ext_mark_uninitialized(&newex);
        err = ext4_ext_insert_extent(handle, inode, path, &newex);
        if (err) {
                /* free data blocks we just allocated */
+               /* not a good idea to call discard here directly,
+                * but otherwise we'd need to call it every free() */
+               ext4_mb_discard_inode_preallocations(inode);
                ext4_free_blocks(handle, inode, ext_pblock(&newex),
-                                       le16_to_cpu(newex.ee_len));
+                                       le16_to_cpu(newex.ee_len), 0);
                goto out2;
        }
 
@@ -2445,6 +2469,7 @@ int ext4_ext_get_blocks(handle_t *handle, struct inode *inode,
 
        /* previous routine could use block we allocated */
        newblock = ext_pblock(&newex);
+       allocated = le16_to_cpu(newex.ee_len);
 outnew:
        __set_bit(BH_New, &bh_result->b_state);
 
@@ -2496,6 +2521,8 @@ void ext4_ext_truncate(struct inode * inode, struct page *page)
        down_write(&EXT4_I(inode)->i_data_sem);
        ext4_ext_invalidate_cache(inode);
 
+       ext4_mb_discard_inode_preallocations(inode);
+
        /*
         * TODO: optimization is possible here.
         * Probably we need not scan at all,
index a06a3b7..bb717cb 100644 (file)
@@ -551,7 +551,7 @@ static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
        return ret;
 failed_out:
        for (i = 0; i <index; i++)
-               ext4_free_blocks(handle, inode, new_blocks[i], 1);
+               ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
        return ret;
 }
 
@@ -650,9 +650,9 @@ failed:
                ext4_journal_forget(handle, branch[i].bh);
        }
        for (i = 0; i <indirect_blks; i++)
-               ext4_free_blocks(handle, inode, new_blocks[i], 1);
+               ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
 
-       ext4_free_blocks(handle, inode, new_blocks[i], num);
+       ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
 
        return err;
 }
@@ -749,9 +749,10 @@ err_out:
        for (i = 1; i <= num; i++) {
                BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
                ext4_journal_forget(handle, where[i].bh);
-               ext4_free_blocks(handle,inode,le32_to_cpu(where[i-1].key),1);
+               ext4_free_blocks(handle, inode,
+                                       le32_to_cpu(where[i-1].key), 1, 0);
        }
-       ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks);
+       ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
 
        return err;
 }
@@ -2052,7 +2053,7 @@ static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
                }
        }
 
-       ext4_free_blocks(handle, inode, block_to_free, count);
+       ext4_free_blocks(handle, inode, block_to_free, count, 0);
 }
 
 /**
@@ -2225,7 +2226,7 @@ static void ext4_free_branches(handle_t *handle, struct inode *inode,
                                ext4_journal_test_restart(handle, inode);
                        }
 
-                       ext4_free_blocks(handle, inode, nr, 1);
+                       ext4_free_blocks(handle, inode, nr, 1, 1);
 
                        if (parent_bh) {
                                /*
diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
new file mode 100644 (file)
index 0000000..76e5fed
--- /dev/null
@@ -0,0 +1,4552 @@
+/*
+ * 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-
+ */
+
+
+/*
+ * mballoc.c contains the multiblocks allocation routines
+ */
+
+#include <linux/time.h>
+#include <linux/fs.h>
+#include <linux/namei.h>
+#include <linux/ext4_jbd2.h>
+#include <linux/ext4_fs.h>
+#include <linux/quotaops.h>
+#include <linux/buffer_head.h>
+#include <linux/module.h>
+#include <linux/swap.h>
+#include <linux/proc_fs.h>
+#include <linux/pagemap.h>
+#include <linux/seq_file.h>
+#include <linux/version.h>
+#include "group.h"
+
+/*
+ * MUSTDO:
+ *   - test ext4_ext_search_left() and ext4_ext_search_right()
+ *   - search for metadata in few groups
+ *
+ * TODO v4:
+ *   - normalization should take into account whether file is still open
+ *   - discard preallocations if no free space left (policy?)
+ *   - don't normalize tails
+ *   - quota
+ *   - reservation for superuser
+ *
+ * TODO v3:
+ *   - bitmap read-ahead (proposed by Oleg Drokin aka green)
+ *   - track min/max extents in each group for better group selection
+ *   - mb_mark_used() may allocate chunk right after splitting buddy
+ *   - tree of groups sorted by number of free blocks
+ *   - error handling
+ */
+
+/*
+ * The allocation request involve request for multiple number of blocks
+ * near to the goal(block) value specified.
+ *
+ * During initialization phase of the allocator we decide to use the group
+ * preallocation or inode preallocation depending on the size file. The
+ * size of the file could be the resulting file size we would have after
+ * allocation or the current file size which ever is larger. If the size is
+ * less that sbi->s_mb_stream_request we select the group
+ * preallocation. The default value of s_mb_stream_request is 16
+ * blocks. This can also be tuned via
+ * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms
+ * of number of blocks.
+ *
+ * The main motivation for having small file use group preallocation is to
+ * ensure that we have small file closer in the disk.
+ *
+ * First stage the allocator looks at the inode prealloc list
+ * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for
+ * this particular inode. The inode prealloc space is represented as:
+ *
+ * pa_lstart -> the logical start block for this prealloc space
+ * pa_pstart -> the physical start block for this prealloc space
+ * pa_len    -> lenght for this prealloc space
+ * pa_free   ->  free space available in this prealloc space
+ *
+ * The inode preallocation space is used looking at the _logical_ start
+ * block. If only the logical file block falls within the range of prealloc
+ * space we will consume the particular prealloc space. This make sure that
+ * that the we have contiguous physical blocks representing the file blocks
+ *
+ * The important thing to be noted in case of inode prealloc space is that
+ * we don't modify the values associated to inode prealloc space except
+ * pa_free.
+ *
+ * If we are not able to find blocks in the inode prealloc space and if we
+ * have the group allocation flag set then we look at the locality group
+ * prealloc space. These are per CPU prealloc list repreasented as
+ *
+ * ext4_sb_info.s_locality_groups[smp_processor_id()]
+ *
+ * The reason for having a per cpu locality group is to reduce the contention
+ * between CPUs. It is possible to get scheduled at this point.
+ *
+ * The locality group prealloc space is used looking at whether we have
+ * enough free space (pa_free) withing the prealloc space.
+ *
+ * If we can't allocate blocks via inode prealloc or/and locality group
+ * prealloc then we look at the buddy cache. The buddy cache is represented
+ * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
+ * mapped to the buddy and bitmap information regarding different
+ * groups. The buddy information is attached to buddy cache inode so that
+ * we can access them through the page cache. The information regarding
+ * each group is loaded via ext4_mb_load_buddy.  The information involve
+ * block bitmap and buddy information. The information are stored in the
+ * inode as:
+ *
+ *  {                        page                        }
+ *  [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
+ *
+ *
+ * one block each for bitmap and buddy information.  So for each group we
+ * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
+ * blocksize) blocks.  So it can have information regarding groups_per_page
+ * which is blocks_per_page/2
+ *
+ * The buddy cache inode is not stored on disk. The inode is thrown
+ * away when the filesystem is unmounted.
+ *
+ * We look for count number of blocks in the buddy cache. If we were able
+ * to locate that many free blocks we return with additional information
+ * regarding rest of the contiguous physical block available
+ *
+ * Before allocating blocks via buddy cache we normalize the request
+ * blocks. This ensure we ask for more blocks that we needed. The extra
+ * blocks that we get after allocation is added to the respective prealloc
+ * list. In case of inode preallocation we follow a list of heuristics
+ * based on file size. This can be found in ext4_mb_normalize_request. If
+ * we are doing a group prealloc we try to normalize the request to
+ * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to
+ * 512 blocks. This can be tuned via
+ * /proc/fs/ext4/<partition/group_prealloc. The value is represented in
+ * terms of number of blocks. If we have mounted the file system with -O
+ * stripe=<value> option the group prealloc request is normalized to the
+ * stripe value (sbi->s_stripe)
+ *
+ * The regular allocator(using the buddy cache) support few tunables.
+ *
+ * /proc/fs/ext4/<partition>/min_to_scan
+ * /proc/fs/ext4/<partition>/max_to_scan
+ * /proc/fs/ext4/<partition>/order2_req
+ *
+ * The regular allocator use buddy scan only if the request len is power of
+ * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
+ * value of s_mb_order2_reqs can be tuned via
+ * /proc/fs/ext4/<partition>/order2_req.  If the request len is equal to
+ * stripe size (sbi->s_stripe), we try to search for contigous block in
+ * stripe size. This should result in better allocation on RAID setup. If
+ * not we search in the specific group using bitmap for best extents. The
+ * tunable min_to_scan and max_to_scan controll the behaviour here.
+ * min_to_scan indicate how long the mballoc __must__ look for a best
+ * extent and max_to_scanindicate how long the mballoc __can__ look for a
+ * best extent in the found extents. Searching for the blocks starts with
+ * the group specified as the goal value in allocation context via
+ * ac_g_ex. Each group is first checked based on the criteria whether it
+ * can used for allocation. ext4_mb_good_group explains how the groups are
+ * checked.
+ *
+ * Both the prealloc space are getting populated as above. So for the first
+ * request we will hit the buddy cache which will result in this prealloc
+ * space getting filled. The prealloc space is then later used for the
+ * subsequent request.
+ */
+
+/*
+ * mballoc operates on the following data:
+ *  - on-disk bitmap
+ *  - in-core buddy (actually includes buddy and bitmap)
+ *  - preallocation descriptors (PAs)
+ *
+ * there are two types of preallocations:
+ *  - inode
+ *    assiged to specific inode and can be used for this inode only.
+ *    it describes part of inode's space preallocated to specific
+ *    physical blocks. any block from that preallocated can be used
+ *    independent. the descriptor just tracks number of blocks left
+ *    unused. so, before taking some block from descriptor, one must
+ *    make sure corresponded logical block isn't allocated yet. this
+ *    also means that freeing any block within descriptor's range
+ *    must discard all preallocated blocks.
+ *  - locality group
+ *    assigned to specific locality group which does not translate to
+ *    permanent set of inodes: inode can join and leave group. space
+ *    from this type of preallocation can be used for any inode. thus
+ *    it's consumed from the beginning to the end.
+ *
+ * relation between them can be expressed as:
+ *    in-core buddy = on-disk bitmap + preallocation descriptors
+ *
+ * this mean blocks mballoc considers used are:
+ *  - allocated blocks (persistent)
+ *  - preallocated blocks (non-persistent)
+ *
+ * consistency in mballoc world means that at any time a block is either
+ * free or used in ALL structures. notice: "any time" should not be read
+ * literally -- time is discrete and delimited by locks.
+ *
+ *  to keep it simple, we don't use block numbers, instead we count number of
+ *  blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
+ *
+ * all operations can be expressed as:
+ *  - init buddy:                      buddy = on-disk + PAs
+ *  - new PA:                          buddy += N; PA = N
+ *  - use inode PA:                    on-disk += N; PA -= N
+ *  - discard inode PA                 buddy -= on-disk - PA; PA = 0
+ *  - use locality group PA            on-disk += N; PA -= N
+ *  - discard locality group PA                buddy -= PA; PA = 0
+ *  note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
+ *        is used in real operation because we can't know actual used
+ *        bits from PA, only from on-disk bitmap
+ *
+ * if we follow this strict logic, then all operations above should be atomic.
+ * given some of them can block, we'd have to use something like semaphores
+ * killing performance on high-end SMP hardware. let's try to relax it using
+ * the following knowledge:
+ *  1) if buddy is referenced, it's already initialized
+ *  2) while block is used in buddy and the buddy is referenced,
+ *     nobody can re-allocate that block
+ *  3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
+ *     bit set and PA claims same block, it's OK. IOW, one can set bit in
+ *     on-disk bitmap if buddy has same bit set or/and PA covers corresponded
+ *     block
+ *
+ * so, now we're building a concurrency table:
+ *  - init buddy vs.
+ *    - new PA
+ *      blocks for PA are allocated in the buddy, buddy must be referenced
+ *      until PA is linked to allocation group to avoid concurrent buddy init
+ *    - use inode PA
+ *      we need to make sure that either on-disk bitmap or PA has uptodate data
+ *      given (3) we care that PA-=N operation doesn't interfere with init
+ *    - discard inode PA
+ *      the simplest way would be to have buddy initialized by the discard
+ *    - use locality group PA
+ *      again PA-=N must be serialized with init
+ *    - discard locality group PA
+ *      the simplest way would be to have buddy initialized by the discard
+ *  - new PA vs.
+ *    - use inode PA
+ *      i_data_sem serializes them
+ *    - discard inode PA
+ *      discard process must wait until PA isn't used by another process
+ *    - use locality group PA
+ *      some mutex should serialize them
+ *    - discard locality group PA
+ *      discard process must wait until PA isn't used by another process
+ *  - use inode PA
+ *    - use inode PA
+ *      i_data_sem or another mutex should serializes them
+ *    - discard inode PA
+ *      discard process must wait until PA isn't used by another process
+ *    - use locality group PA
+ *      nothing wrong here -- they're different PAs covering different blocks
+ *    - discard locality group PA
+ *      discard process must wait until PA isn't used by another process
+ *
+ * now we're ready to make few consequences:
+ *  - PA is referenced and while it is no discard is possible
+ *  - PA is referenced until block isn't marked in on-disk bitmap
+ *  - PA changes only after on-disk bitmap
+ *  - discard must not compete with init. either init is done before
+ *    any discard or they're serialized somehow
+ *  - buddy init as sum of on-disk bitmap and PAs is done atomically
+ *
+ * a special case when we've used PA to emptiness. no need to modify buddy
+ * in this case, but we should care about concurrent init
+ *
+ */
+
+ /*
+ * Logic in few words:
+ *
+ *  - allocation:
+ *    load group
+ *    find blocks
+ *    mark bits in on-disk bitmap
+ *    release group
+ *
+ *  - use preallocation:
+ *    find proper PA (per-inode or group)
+ *    load group
+ *    mark bits in on-disk bitmap
+ *    release group
+ *    release PA
+ *
+ *  - free:
+ *    load group
+ *    mark bits in on-disk bitmap
+ *    release group
+ *
+ *  - discard preallocations in group:
+ *    mark PAs deleted
+ *    move them onto local list
+ *    load on-disk bitmap
+ *    load group
+ *    remove PA from object (inode or locality group)
+ *    mark free blocks in-core
+ *
+ *  - discard inode's preallocations:
+ */
+
+/*
+ * Locking rules
+ *
+ * Locks:
+ *  - bitlock on a group       (group)
+ *  - object (inode/locality)  (object)
+ *  - per-pa lock              (pa)
+ *
+ * Paths:
+ *  - new pa
+ *    object
+ *    group
+ *
+ *  - find and use pa:
+ *    pa
+ *
+ *  - release consumed pa:
+ *    pa
+ *    group
+ *    object
+ *
+ *  - generate in-core bitmap:
+ *    group
+ *        pa
+ *
+ *  - discard all for given object (inode, locality group):
+ *    object
+ *        pa
+ *    group
+ *
+ *  - discard all for given group:
+ *    group
+ *        pa
+ *    group
+ *        object
+ *
+ */
+
+/*
+ * with AGGRESSIVE_CHECK allocator runs consistency checks over
+ * structures. these checks slow things down a lot
+ */
+#define AGGRESSIVE_CHECK__
+
+/*
+ * with DOUBLE_CHECK defined mballoc creates persistent in-core
+ * bitmaps, maintains and uses them to check for double allocations
+ */
+#define DOUBLE_CHECK__
+
+/*
+ */
+#define MB_DEBUG__
+#ifdef MB_DEBUG
+#define mb_debug(fmt, a...)    printk(fmt, ##a)
+#else
+#define mb_debug(fmt, a...)
+#endif
+
+/*
+ * with EXT4_MB_HISTORY mballoc stores last N allocations in memory
+ * and you can monitor it in /proc/fs/ext4/<dev>/mb_history
+ */
+#define EXT4_MB_HISTORY
+#define EXT4_MB_HISTORY_ALLOC          1       /* allocation */
+#define EXT4_MB_HISTORY_PREALLOC       2       /* preallocated blocks used */
+#define EXT4_MB_HISTORY_DISCARD                4       /* preallocation discarded */
+#define EXT4_MB_HISTORY_FREE           8       /* free */
+
+#define EXT4_MB_HISTORY_DEFAULT                (EXT4_MB_HISTORY_ALLOC | \
+                                        EXT4_MB_HISTORY_PREALLOC)
+
+/*
+ * How long mballoc can look for a best extent (in found extents)
+ */
+#define MB_DEFAULT_MAX_TO_SCAN         200
+
+/*
+ * How long mballoc must look for a best extent
+ */
+#define MB_DEFAULT_MIN_TO_SCAN         10
+
+/*
+ * How many groups mballoc will scan looking for the best chunk
+ */
+#define MB_DEFAULT_MAX_GROUPS_TO_SCAN  5
+
+/*
+ * with 'ext4_mb_stats' allocator will collect stats that will be
+ * shown at umount. The collecting costs though!
+ */
+#define MB_DEFAULT_STATS               1
+
+/*
+ * files smaller than MB_DEFAULT_STREAM_THRESHOLD are served
+ * by the stream allocator, which purpose is to pack requests
+ * as close each to other as possible to produce smooth I/O traffic
+ * We use locality group prealloc space for stream request.
+ * We can tune the same via /proc/fs/ext4/<parition>/stream_req
+ */
+#define MB_DEFAULT_STREAM_THRESHOLD    16      /* 64K */
+
+/*
+ * for which requests use 2^N search using buddies
+ */
+#define MB_DEFAULT_ORDER2_REQS         2
+
+/*
+ * default group prealloc size 512 blocks
+ */
+#define MB_DEFAULT_GROUP_PREALLOC      512
+
+static struct kmem_cache *ext4_pspace_cachep;
+
+#ifdef EXT4_BB_MAX_BLOCKS
+#undef EXT4_BB_MAX_BLOCKS
+#endif
+#define EXT4_BB_MAX_BLOCKS     30
+
+struct ext4_free_metadata {
+       ext4_group_t group;
+       unsigned short num;
+       ext4_grpblk_t  blocks[EXT4_BB_MAX_BLOCKS];
+       struct list_head list;
+};
+
+struct ext4_group_info {
+       unsigned long   bb_state;
+       unsigned long   bb_tid;
+       struct ext4_free_metadata *bb_md_cur;
+       unsigned short  bb_first_free;
+       unsigned short  bb_free;
+       unsigned short  bb_fragments;
+       struct          list_head bb_prealloc_list;
+#ifdef DOUBLE_CHECK
+       void            *bb_bitmap;
+#endif
+       unsigned short  bb_counters[];
+};
+
+#define EXT4_GROUP_INFO_NEED_INIT_BIT  0
+#define EXT4_GROUP_INFO_LOCKED_BIT     1
+
+#define EXT4_MB_GRP_NEED_INIT(grp)     \
+       (test_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &((grp)->bb_state)))
+
+
+struct ext4_prealloc_space {
+       struct list_head        pa_inode_list;
+       struct list_head        pa_group_list;
+       union {
+               struct list_head pa_tmp_list;
+               struct rcu_head pa_rcu;
+       } u;
+       spinlock_t              pa_lock;
+       atomic_t                pa_count;
+       unsigned                pa_deleted;
+       ext4_fsblk_t            pa_pstart;      /* phys. block */
+       ext4_lblk_t             pa_lstart;      /* log. block */
+       unsigned short          pa_len;         /* len of preallocated chunk */
+       unsigned short          pa_free;        /* how many blocks are free */
+       unsigned short          pa_linear;      /* consumed in one direction
+                                                * strictly, for grp prealloc */
+       spinlock_t              *pa_obj_lock;
+       struct inode            *pa_inode;      /* hack, for history only */
+};
+
+
+struct ext4_free_extent {
+       ext4_lblk_t fe_logical;
+       ext4_grpblk_t fe_start;
+       ext4_group_t fe_group;
+       int fe_len;
+};
+
+/*
+ * Locality group:
+ *   we try to group all related changes together
+ *   so that writeback can flush/allocate them together as well
+ */
+struct ext4_locality_group {
+       /* for allocator */
+       struct mutex            lg_mutex;       /* to serialize allocates */
+       struct list_head        lg_prealloc_list;/* list of preallocations */
+       spinlock_t              lg_prealloc_lock;
+};
+
+struct ext4_allocation_context {
+       struct inode *ac_inode;
+       struct super_block *ac_sb;
+
+       /* original request */
+       struct ext4_free_extent ac_o_ex;
+
+       /* goal request (after normalization) */
+       struct ext4_free_extent ac_g_ex;
+
+       /* the best found extent */
+       struct ext4_free_extent ac_b_ex;
+
+       /* copy of the bext found extent taken before preallocation efforts */
+       struct ext4_free_extent ac_f_ex;
+
+       /* number of iterations done. we have to track to limit searching */
+       unsigned long ac_ex_scanned;
+       __u16 ac_groups_scanned;
+       __u16 ac_found;
+       __u16 ac_tail;
+       __u16 ac_buddy;
+       __u16 ac_flags;         /* allocation hints */
+       __u8 ac_status;
+       __u8 ac_criteria;
+       __u8 ac_repeats;
+       __u8 ac_2order;         /* if request is to allocate 2^N blocks and
+                                * N > 0, the field stores N, otherwise 0 */
+       __u8 ac_op;             /* operation, for history only */
+       struct page *ac_bitmap_page;
+       struct page *ac_buddy_page;
+       struct ext4_prealloc_space *ac_pa;
+       struct ext4_locality_group *ac_lg;
+};
+
+#define AC_STATUS_CONTINUE     1
+#define AC_STATUS_FOUND                2
+#define AC_STATUS_BREAK                3
+
+struct ext4_mb_history {
+       struct ext4_free_extent orig;   /* orig allocation */
+       struct ext4_free_extent goal;   /* goal allocation */
+       struct ext4_free_extent result; /* result allocation */
+       unsigned pid;
+       unsigned ino;
+       __u16 found;    /* how many extents have been found */
+       __u16 groups;   /* how many groups have been scanned */
+       __u16 tail;     /* what tail broke some buddy */
+       __u16 buddy;    /* buddy the tail ^^^ broke */
+       __u16 flags;
+       __u8 cr:3;      /* which phase the result extent was found at */
+       __u8 op:4;
+       __u8 merged:1;
+};
+
+struct ext4_buddy {
+       struct page *bd_buddy_page;
+       void *bd_buddy;
+       struct page *bd_bitmap_page;
+       void *bd_bitmap;
+       struct ext4_group_info *bd_info;
+       struct super_block *bd_sb;
+       __u16 bd_blkbits;
+       ext4_group_t bd_group;
+};
+#define EXT4_MB_BITMAP(e4b)    ((e4b)->bd_bitmap)
+#define EXT4_MB_BUDDY(e4b)     ((e4b)->bd_buddy)
+
+#ifndef EXT4_MB_HISTORY
+static inline void ext4_mb_store_history(struct ext4_allocation_context *ac)
+{
+       return;
+}
+#else
+static void ext4_mb_store_history(struct ext4_allocation_context *ac);
+#endif
+
+#define in_range(b, first, len)        ((b) >= (first) && (b) <= (first) + (len) - 1)
+
+static struct proc_dir_entry *proc_root_ext4;
+struct buffer_head *read_block_bitmap(struct super_block *, ext4_group_t);
+ext4_fsblk_t ext4_new_blocks_old(handle_t *handle, struct inode *inode,
+                       ext4_fsblk_t goal, unsigned long *count, int *errp);
+
+static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
+                                       ext4_group_t group);
+static void ext4_mb_poll_new_transaction(struct super_block *, handle_t *);
+static void ext4_mb_free_committed_blocks(struct super_block *);
+static void ext4_mb_return_to_preallocation(struct inode *inode,
+                                       struct ext4_buddy *e4b, sector_t block,
+                                       int count);
+static void ext4_mb_put_pa(struct ext4_allocation_context *,
+                       struct super_block *, struct ext4_prealloc_space *pa);
+static int ext4_mb_init_per_dev_proc(struct super_block *sb);
+static int ext4_mb_destroy_per_dev_proc(struct super_block *sb);
+
+
+static inline void ext4_lock_group(struct super_block *sb, ext4_group_t group)
+{
+       struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
+
+       bit_spin_lock(EXT4_GROUP_INFO_LOCKED_BIT, &(grinfo->bb_state));
+}
+
+static inline void ext4_unlock_group(struct super_block *sb,
+                                       ext4_group_t group)
+{
+       struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
+
+       bit_spin_unlock(EXT4_GROUP_INFO_LOCKED_BIT, &(grinfo->bb_state));
+}
+
+static inline int ext4_is_group_locked(struct super_block *sb,
+                                       ext4_group_t group)
+{
+       struct ext4_group_info *grinfo = ext4_get_group_info(sb, group);
+
+       return bit_spin_is_locked(EXT4_GROUP_INFO_LOCKED_BIT,
+                                               &(grinfo->bb_state));
+}
+
+static ext4_fsblk_t ext4_grp_offs_to_block(struct super_block *sb,
+                                       struct ext4_free_extent *fex)
+{
+       ext4_fsblk_t block;
+
+       block = (ext4_fsblk_t) fex->fe_group * EXT4_BLOCKS_PER_GROUP(sb)
+                       + fex->fe_start
+                       + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
+       return block;
+}
+
+#if BITS_PER_LONG == 64
+#define mb_correct_addr_and_bit(bit, addr)             \
+{                                                      \
+       bit += ((unsigned long) addr & 7UL) << 3;       \
+       addr = (void *) ((unsigned long) addr & ~7UL);  \
+}
+#elif BITS_PER_LONG == 32
+#define mb_correct_addr_and_bit(bit, addr)             \
+{                                                      \
+       bit += ((unsigned long) addr & 3UL) << 3;       \
+       addr = (void *) ((unsigned long) addr & ~3UL);  \
+}
+#else
+#error "how many bits you are?!"
+#endif
+
+static inline int mb_test_bit(int bit, void *addr)
+{
+       /*
+        * ext4_test_bit on architecture like powerpc
+        * needs unsigned long aligned address
+        */
+       mb_correct_addr_and_bit(bit, addr);
+       return ext4_test_bit(bit, addr);
+}
+
+static inline void mb_set_bit(int bit, void *addr)
+{
+       mb_correct_addr_and_bit(bit, addr);
+       ext4_set_bit(bit, addr);
+}
+
+static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr)
+{
+       mb_correct_addr_and_bit(bit, addr);
+       ext4_set_bit_atomic(lock, bit, addr);
+}
+
+static inline void mb_clear_bit(int bit, void *addr)
+{
+       mb_correct_addr_and_bit(bit, addr);
+       ext4_clear_bit(bit, addr);
+}
+
+static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr)
+{
+       mb_correct_addr_and_bit(bit, addr);
+       ext4_clear_bit_atomic(lock, bit, addr);
+}
+
+static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
+{
+       char *bb;
+
+       /* FIXME!! is this needed */
+       BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
+       BUG_ON(max == NULL);
+
+       if (order > e4b->bd_blkbits + 1) {
+               *max = 0;
+               return NULL;
+       }
+
+       /* at order 0 we see each particular block */
+       *max = 1 << (e4b->bd_blkbits + 3);
+       if (order == 0)
+               return EXT4_MB_BITMAP(e4b);
+
+       bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
+       *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
+
+       return bb;
+}
+
+#ifdef DOUBLE_CHECK
+static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
+                          int first, int count)
+{
+       int i;
+       struct super_block *sb = e4b->bd_sb;
+
+       if (unlikely(e4b->bd_info->bb_bitmap == NULL))
+               return;
+       BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
+       for (i = 0; i < count; i++) {
+               if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
+                       ext4_fsblk_t blocknr;
+                       blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
+                       blocknr += first + i;
+                       blocknr +=
+                           le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
+
+                       ext4_error(sb, __FUNCTION__, "double-free of inode"
+                                  " %lu's block %llu(bit %u in group %lu)\n",
+                                  inode ? inode->i_ino : 0, blocknr,
+                                  first + i, e4b->bd_group);
+               }
+               mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
+       }
+}
+
+static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
+{
+       int i;
+
+       if (unlikely(e4b->bd_info->bb_bitmap == NULL))
+               return;
+       BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
+       for (i = 0; i < count; i++) {
+               BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
+               mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
+       }
+}
+
+static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
+{
+       if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
+               unsigned char *b1, *b2;
+               int i;
+               b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
+               b2 = (unsigned char *) bitmap;
+               for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
+                       if (b1[i] != b2[i]) {
+                               printk("corruption in group %lu at byte %u(%u):"
+                                      " %x in copy != %x on disk/prealloc\n",
+                                       e4b->bd_group, i, i * 8, b1[i], b2[i]);
+                               BUG();
+                       }
+               }
+       }
+}
+
+#else
+static inline void mb_free_blocks_double(struct inode *inode,
+                               struct ext4_buddy *e4b, int first, int count)
+{
+       return;
+}
+static inline void mb_mark_used_double(struct ext4_buddy *e4b,
+                                               int first, int count)
+{
+       return;
+}
+static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
+{
+       return;
+}
+#endif
+
+#ifdef AGGRESSIVE_CHECK
+
+#define MB_CHECK_ASSERT(assert)                                                \
+do {                                                                   \
+       if (!(assert)) {                                                \
+               printk(KERN_EMERG                                       \
+                       "Assertion failure in %s() at %s:%d: \"%s\"\n", \
+                       function, file, line, # assert);                \
+               BUG();                                                  \
+       }                                                               \
+} while (0)
+
+static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
+                               const char *function, int line)
+{
+       struct super_block *sb = e4b->bd_sb;
+       int order = e4b->bd_blkbits + 1;
+       int max;
+       int max2;
+       int i;
+       int j;
+       int k;
+       int count;
+       struct ext4_group_info *grp;
+       int fragments = 0;
+       int fstart;
+       struct list_head *cur;
+       void *buddy;
+       void *buddy2;
+
+       if (!test_opt(sb, MBALLOC))
+               return 0;
+
+       {
+               static int mb_check_counter;
+               if (mb_check_counter++ % 100 != 0)
+                       return 0;
+       }
+
+       while (order > 1) {
+               buddy = mb_find_buddy(e4b, order, &max);
+               MB_CHECK_ASSERT(buddy);
+               buddy2 = mb_find_buddy(e4b, order - 1, &max2);
+               MB_CHECK_ASSERT(buddy2);
+               MB_CHECK_ASSERT(buddy != buddy2);
+               MB_CHECK_ASSERT(max * 2 == max2);
+
+               count = 0;
+               for (i = 0; i < max; i++) {
+
+                       if (mb_test_bit(i, buddy)) {
+                               /* only single bit in buddy2 may be 1 */
+                               if (!mb_test_bit(i << 1, buddy2)) {
+                                       MB_CHECK_ASSERT(
+                                               mb_test_bit((i<<1)+1, buddy2));
+                               } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
+                                       MB_CHECK_ASSERT(
+                                               mb_test_bit(i << 1, buddy2));
+                               }
+                               continue;
+                       }
+
+                       /* both bits in buddy2 must be 0 */
+                       MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
+                       MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
+
+                       for (j = 0; j < (1 << order); j++) {
+                               k = (i * (1 << order)) + j;
+                               MB_CHECK_ASSERT(
+                                       !mb_test_bit(k, EXT4_MB_BITMAP(e4b)));
+                       }
+                       count++;
+               }
+               MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
+               order--;
+       }
+
+       fstart = -1;
+       buddy = mb_find_buddy(e4b, 0, &max);
+       for (i = 0; i < max; i++) {
+               if (!mb_test_bit(i, buddy)) {
+                       MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
+                       if (fstart == -1) {
+                               fragments++;
+                               fstart = i;
+                       }
+                       continue;
+               }
+               fstart = -1;
+               /* check used bits only */
+               for (j = 0; j < e4b->bd_blkbits + 1; j++) {
+                       buddy2 = mb_find_buddy(e4b, j, &max2);
+                       k = i >> j;
+                       MB_CHECK_ASSERT(k < max2);
+                       MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
+               }
+       }
+       MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
+       MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
+
+       grp = ext4_get_group_info(sb, e4b->bd_group);
+       buddy = mb_find_buddy(e4b, 0, &max);
+       list_for_each(cur, &grp->bb_prealloc_list) {
+               ext4_group_t groupnr;
+               struct ext4_prealloc_space *pa;
+               pa = list_entry(cur, struct ext4_prealloc_space, group_list);
+               ext4_get_group_no_and_offset(sb, pa->pstart, &groupnr, &k);
+               MB_CHECK_ASSERT(groupnr == e4b->bd_group);
+               for (i = 0; i < pa->len; i++)
+                       MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
+       }
+       return 0;
+}
+#undef MB_CHECK_ASSERT
+#define mb_check_buddy(e4b) __mb_check_buddy(e4b,      \
+                                       __FILE__, __FUNCTION__, __LINE__)
+#else
+#define mb_check_buddy(e4b)
+#endif
+
+/* FIXME!! need more doc */
+static void ext4_mb_mark_free_simple(struct super_block *sb,
+                               void *buddy, unsigned first, int len,
+                                       struct ext4_group_info *grp)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       unsigned short min;
+       unsigned short max;
+       unsigned short chunk;
+       unsigned short border;
+
+       BUG_ON(len >= EXT4_BLOCKS_PER_GROUP(sb));
+
+       border = 2 << sb->s_blocksize_bits;
+
+       while (len > 0) {
+               /* find how many blocks can be covered since this position */
+               max = ffs(first | border) - 1;
+
+               /* find how many blocks of power 2 we need to mark */
+               min = fls(len) - 1;
+
+               if (max < min)
+                       min = max;
+               chunk = 1 << min;
+
+               /* mark multiblock chunks only */
+               grp->bb_counters[min]++;
+               if (min > 0)
+                       mb_clear_bit(first >> min,
+                                    buddy + sbi->s_mb_offsets[min]);
+
+               len -= chunk;
+               first += chunk;
+       }
+}
+
+static void ext4_mb_generate_buddy(struct super_block *sb,
+                               void *buddy, void *bitmap, ext4_group_t group)
+{
+       struct ext4_group_info *grp = ext4_get_group_info(sb, group);
+       unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
+       unsigned short i = 0;
+       unsigned short first;
+       unsigned short len;
+       unsigned free = 0;
+       unsigned fragments = 0;
+       unsigned long long period = get_cycles();
+
+       /* initialize buddy from bitmap which is aggregation
+        * of on-disk bitmap and preallocations */
+       i = ext4_find_next_zero_bit(bitmap, max, 0);
+       grp->bb_first_free = i;
+       while (i < max) {
+               fragments++;
+               first = i;
+               i = ext4_find_next_bit(bitmap, max, i);
+               len = i - first;
+               free += len;
+               if (len > 1)
+                       ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
+               else
+                       grp->bb_counters[0]++;
+               if (i < max)
+                       i = ext4_find_next_zero_bit(bitmap, max, i);
+       }
+       grp->bb_fragments = fragments;
+
+       if (free != grp->bb_free) {
+               printk(KERN_DEBUG
+                       "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
+                       group, free, grp->bb_free);
+               grp->bb_free = free;
+       }
+
+       clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
+
+       period = get_cycles() - period;
+       spin_lock(&EXT4_SB(sb)->s_bal_lock);
+       EXT4_SB(sb)->s_mb_buddies_generated++;
+       EXT4_SB(sb)->s_mb_generation_time += period;
+       spin_unlock(&EXT4_SB(sb)->s_bal_lock);
+}
+
+/* The buddy information is attached the buddy cache inode
+ * for convenience. The information regarding each group
+ * is loaded via ext4_mb_load_buddy. The information involve
+ * block bitmap and buddy information. The information are
+ * stored in the inode as
+ *
+ * {                        page                        }
+ * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]...
+ *
+ *
+ * one block each for bitmap and buddy information.
+ * So for each group we take up 2 blocks. A page can
+ * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize)  blocks.
+ * So it can have information regarding groups_per_page which
+ * is blocks_per_page/2
+ */
+
+static int ext4_mb_init_cache(struct page *page, char *incore)
+{
+       int blocksize;
+       int blocks_per_page;
+       int groups_per_page;
+       int err = 0;
+       int i;
+       ext4_group_t first_group;
+       int first_block;
+       struct super_block *sb;
+       struct buffer_head *bhs;
+       struct buffer_head **bh;
+       struct inode *inode;
+       char *data;
+       char *bitmap;
+
+       mb_debug("init page %lu\n", page->index);
+
+       inode = page->mapping->host;
+       sb = inode->i_sb;
+       blocksize = 1 << inode->i_blkbits;
+       blocks_per_page = PAGE_CACHE_SIZE / blocksize;
+
+       groups_per_page = blocks_per_page >> 1;
+       if (groups_per_page == 0)
+               groups_per_page = 1;
+
+       /* allocate buffer_heads to read bitmaps */
+       if (groups_per_page > 1) {
+               err = -ENOMEM;
+               i = sizeof(struct buffer_head *) * groups_per_page;
+               bh = kzalloc(i, GFP_NOFS);
+               if (bh == NULL)
+                       goto out;
+       } else
+               bh = &bhs;
+
+       first_group = page->index * blocks_per_page / 2;
+
+       /* read all groups the page covers into the cache */
+       for (i = 0; i < groups_per_page; i++) {
+               struct ext4_group_desc *desc;
+
+               if (first_group + i >= EXT4_SB(sb)->s_groups_count)
+                       break;
+
+               err = -EIO;
+               desc = ext4_get_group_desc(sb, first_group + i, NULL);
+               if (desc == NULL)
+                       goto out;
+
+               err = -ENOMEM;
+               bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc));
+               if (bh[i] == NULL)
+                       goto out;
+
+               if (bh_uptodate_or_lock(bh[i]))
+                       continue;
+
+               if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
+                       ext4_init_block_bitmap(sb, bh[i],
+                                               first_group + i, desc);
+                       set_buffer_uptodate(bh[i]);
+                       unlock_buffer(bh[i]);
+                       continue;
+               }
+               get_bh(bh[i]);
+               bh[i]->b_end_io = end_buffer_read_sync;
+               submit_bh(READ, bh[i]);
+               mb_debug("read bitmap for group %lu\n", first_group + i);
+       }
+
+       /* wait for I/O completion */
+       for (i = 0; i < groups_per_page && bh[i]; i++)
+               wait_on_buffer(bh[i]);
+
+       err = -EIO;
+       for (i = 0; i < groups_per_page && bh[i]; i++)
+               if (!buffer_uptodate(bh[i]))
+                       goto out;
+
+       first_block = page->index * blocks_per_page;
+       for (i = 0; i < blocks_per_page; i++) {
+               int group;
+               struct ext4_group_info *grinfo;
+
+               group = (first_block + i) >> 1;
+               if (group >= EXT4_SB(sb)->s_groups_count)
+                       break;
+
+               /*
+                * data carry information regarding this
+                * particular group in the format specified
+                * above
+                *
+                */
+               data = page_address(page) + (i * blocksize);
+               bitmap = bh[group - first_group]->b_data;
+
+               /*
+                * We place the buddy block and bitmap block
+                * close together
+                */
+               if ((first_block + i) & 1) {
+                       /* this is block of buddy */
+                       BUG_ON(incore == NULL);
+                       mb_debug("put buddy for group %u in page %lu/%x\n",
+                               group, page->index, i * blocksize);
+                       memset(data, 0xff, blocksize);
+                       grinfo = ext4_get_group_info(sb, group);
+                       grinfo->bb_fragments = 0;
+                       memset(grinfo->bb_counters, 0,
+                              sizeof(unsigned short)*(sb->s_blocksize_bits+2));
+                       /*
+                        * incore got set to the group block bitmap below
+                        */
+                       ext4_mb_generate_buddy(sb, data, incore, group);
+                       incore = NULL;
+               } else {
+                       /* this is block of bitmap */
+                       BUG_ON(incore != NULL);
+                       mb_debug("put bitmap for group %u in page %lu/%x\n",
+                               group, page->index, i * blocksize);
+
+                       /* see comments in ext4_mb_put_pa() */
+                       ext4_lock_group(sb, group);
+                       memcpy(data, bitmap, blocksize);
+
+                       /* mark all preallocated blks used in in-core bitmap */
+                       ext4_mb_generate_from_pa(sb, data, group);
+                       ext4_unlock_group(sb, group);
+
+                       /* set incore so that the buddy information can be
+                        * generated using this
+                        */
+                       incore = data;
+               }
+       }
+       SetPageUptodate(page);
+
+out:
+       if (bh) {
+               for (i = 0; i < groups_per_page && bh[i]; i++)
+                       brelse(bh[i]);
+               if (bh != &bhs)
+                       kfree(bh);
+       }
+       return err;
+}
+
+static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
+               struct ext4_buddy *e4b)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       struct inode *inode = sbi->s_buddy_cache;
+       int blocks_per_page;
+       int block;
+       int pnum;
+       int poff;
+       struct page *page;
+
+       mb_debug("load group %lu\n", group);
+
+       blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
+
+       e4b->bd_blkbits = sb->s_blocksize_bits;
+       e4b->bd_info = ext4_get_group_info(sb, group);
+       e4b->bd_sb = sb;
+       e4b->bd_group = group;
+       e4b->bd_buddy_page = NULL;
+       e4b->bd_bitmap_page = NULL;
+
+       /*
+        * the buddy cache inode stores the block bitmap
+        * and buddy information in consecutive blocks.
+        * So for each group we need two blocks.
+        */
+       block = group * 2;
+       pnum = block / blocks_per_page;
+       poff = block % blocks_per_page;
+
+       /* we could use find_or_create_page(), but it locks page
+        * what we'd like to avoid in fast path ... */
+       page = find_get_page(inode->i_mapping, pnum);
+       if (page == NULL || !PageUptodate(page)) {
+               if (page)
+                       page_cache_release(page);
+               page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
+               if (page) {
+                       BUG_ON(page->mapping != inode->i_mapping);
+                       if (!PageUptodate(page)) {
+                               ext4_mb_init_cache(page, NULL);
+                               mb_cmp_bitmaps(e4b, page_address(page) +
+                                              (poff * sb->s_blocksize));
+                       }
+                       unlock_page(page);
+               }
+       }
+       if (page == NULL || !PageUptodate(page))
+               goto err;
+       e4b->bd_bitmap_page = page;
+       e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
+       mark_page_accessed(page);
+
+       block++;
+       pnum = block / blocks_per_page;
+       poff = block % blocks_per_page;
+
+       page = find_get_page(inode->i_mapping, pnum);
+       if (page == NULL || !PageUptodate(page)) {
+               if (page)
+                       page_cache_release(page);
+               page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
+               if (page) {
+                       BUG_ON(page->mapping != inode->i_mapping);
+                       if (!PageUptodate(page))
+                               ext4_mb_init_cache(page, e4b->bd_bitmap);
+
+                       unlock_page(page);
+               }
+       }
+       if (page == NULL || !PageUptodate(page))
+               goto err;
+       e4b->bd_buddy_page = page;
+       e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
+       mark_page_accessed(page);
+
+       BUG_ON(e4b->bd_bitmap_page == NULL);
+       BUG_ON(e4b->bd_buddy_page == NULL);
+
+       return 0;
+
+err:
+       if (e4b->bd_bitmap_page)
+               page_cache_release(e4b->bd_bitmap_page);
+       if (e4b->bd_buddy_page)
+               page_cache_release(e4b->bd_buddy_page);
+       e4b->bd_buddy = NULL;
+       e4b->bd_bitmap = NULL;
+       return -EIO;
+}
+
+static void ext4_mb_release_desc(struct ext4_buddy *e4b)
+{
+       if (e4b->bd_bitmap_page)
+               page_cache_release(e4b->bd_bitmap_page);
+       if (e4b->bd_buddy_page)
+               page_cache_release(e4b->bd_buddy_page);
+}
+
+
+static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
+{
+       int order = 1;
+       void *bb;
+
+       BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b));
+       BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
+
+       bb = EXT4_MB_BUDDY(e4b);
+       while (order <= e4b->bd_blkbits + 1) {
+               block = block >> 1;
+               if (!mb_test_bit(block, bb)) {
+                       /* this block is part of buddy of order 'order' */
+                       return order;
+               }
+               bb += 1 << (e4b->bd_blkbits - order);
+               order++;
+       }
+       return 0;
+}
+
+static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len)
+{
+       __u32 *addr;
+
+       len = cur + len;
+       while (cur < len) {
+               if ((cur & 31) == 0 && (len - cur) >= 32) {
+                       /* fast path: clear whole word at once */
+                       addr = bm + (cur >> 3);
+                       *addr = 0;
+                       cur += 32;
+                       continue;
+               }
+               mb_clear_bit_atomic(lock, cur, bm);
+               cur++;
+       }
+}
+
+static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len)
+{
+       __u32 *addr;
+
+       len = cur + len;
+       while (cur < len) {
+               if ((cur & 31) == 0 && (len - cur) >= 32) {
+                       /* fast path: set whole word at once */
+                       addr = bm + (cur >> 3);
+                       *addr = 0xffffffff;
+                       cur += 32;
+                       continue;
+               }
+               mb_set_bit_atomic(lock, cur, bm);
+               cur++;
+       }
+}
+
+static int mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
+                         int first, int count)
+{
+       int block = 0;
+       int max = 0;
+       int order;
+       void *buddy;
+       void *buddy2;
+       struct super_block *sb = e4b->bd_sb;
+
+       BUG_ON(first + count > (sb->s_blocksize << 3));
+       BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group));
+       mb_check_buddy(e4b);
+       mb_free_blocks_double(inode, e4b, first, count);
+
+       e4b->bd_info->bb_free += count;
+       if (first < e4b->bd_info->bb_first_free)
+               e4b->bd_info->bb_first_free = first;
+
+       /* let's maintain fragments counter */
+       if (first != 0)
+               block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b));
+       if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
+               max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b));
+       if (block && max)
+               e4b->bd_info->bb_fragments--;
+       else if (!block && !max)
+               e4b->bd_info->bb_fragments++;
+
+       /* let's maintain buddy itself */
+       while (count-- > 0) {
+               block = first++;
+               order = 0;
+
+               if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) {
+                       ext4_fsblk_t blocknr;
+                       blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb);
+                       blocknr += block;
+                       blocknr +=
+                           le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
+
+                       ext4_error(sb, __FUNCTION__, "double-free of inode"
+                                  " %lu's block %llu(bit %u in group %lu)\n",
+                                  inode ? inode->i_ino : 0, blocknr, block,
+                                  e4b->bd_group);
+               }
+               mb_clear_bit(block, EXT4_MB_BITMAP(e4b));
+               e4b->bd_info->bb_counters[order]++;
+
+               /* start of the buddy */
+               buddy = mb_find_buddy(e4b, order, &max);
+
+               do {
+                       block &= ~1UL;
+                       if (mb_test_bit(block, buddy) ||
+                                       mb_test_bit(block + 1, buddy))
+                               break;
+
+                       /* both the buddies are free, try to coalesce them */
+                       buddy2 = mb_find_buddy(e4b, order + 1, &max);
+
+                       if (!buddy2)
+                               break;
+
+                       if (order > 0) {
+                               /* for special purposes, we don't set
+                                * free bits in bitmap */
+                               mb_set_bit(block, buddy);
+                               mb_set_bit(block + 1, buddy);
+                       }
+                       e4b->bd_info->bb_counters[order]--;
+                       e4b->bd_info->bb_counters[order]--;
+
+                       block = block >> 1;
+                       order++;
+                       e4b->bd_info->bb_counters[order]++;
+
+                       mb_clear_bit(block, buddy2);
+                       buddy = buddy2;
+               } while (1);
+       }
+       mb_check_buddy(e4b);
+
+       return 0;
+}
+
+static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
+                               int needed, struct ext4_free_extent *ex)
+{
+       int next = block;
+       int max;
+       int ord;
+       void *buddy;
+
+       BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
+       BUG_ON(ex == NULL);
+
+       buddy = mb_find_buddy(e4b, order, &max);
+       BUG_ON(buddy == NULL);
+       BUG_ON(block >= max);
+       if (mb_test_bit(block, buddy)) {
+               ex->fe_len = 0;
+               ex->fe_start = 0;
+               ex->fe_group = 0;
+               return 0;
+       }
+
+       /* FIXME dorp order completely ? */
+       if (likely(order == 0)) {
+               /* find actual order */
+               order = mb_find_order_for_block(e4b, block);
+               block = block >> order;
+       }
+
+       ex->fe_len = 1 << order;
+       ex->fe_start = block << order;
+       ex->fe_group = e4b->bd_group;
+
+       /* calc difference from given start */
+       next = next - ex->fe_start;
+       ex->fe_len -= next;
+       ex->fe_start += next;
+
+       while (needed > ex->fe_len &&
+              (buddy = mb_find_buddy(e4b, order, &max))) {
+
+               if (block + 1 >= max)
+                       break;
+
+               next = (block + 1) * (1 << order);
+               if (mb_test_bit(next, EXT4_MB_BITMAP(e4b)))
+                       break;
+
+               ord = mb_find_order_for_block(e4b, next);
+
+               order = ord;
+               block = next >> order;
+               ex->fe_len += 1 << order;
+       }
+
+       BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
+       return ex->fe_len;
+}
+
+static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
+{
+       int ord;
+       int mlen = 0;
+       int max = 0;
+       int cur;
+       int start = ex->fe_start;
+       int len = ex->fe_len;
+       unsigned ret = 0;
+       int len0 = len;
+       void *buddy;
+
+       BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
+       BUG_ON(e4b->bd_group != ex->fe_group);
+       BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group));
+       mb_check_buddy(e4b);
+       mb_mark_used_double(e4b, start, len);
+
+       e4b->bd_info->bb_free -= len;
+       if (e4b->bd_info->bb_first_free == start)
+               e4b->bd_info->bb_first_free += len;
+
+       /* let's maintain fragments counter */
+       if (start != 0)
+               mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b));
+       if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
+               max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b));
+       if (mlen && max)
+               e4b->bd_info->bb_fragments++;
+       else if (!mlen && !max)
+               e4b->bd_info->bb_fragments--;
+
+       /* let's maintain buddy itself */
+       while (len) {
+               ord = mb_find_order_for_block(e4b, start);
+
+               if (((start >> ord) << ord) == start && len >= (1 << ord)) {
+                       /* the whole chunk may be allocated at once! */
+                       mlen = 1 << ord;
+                       buddy = mb_find_buddy(e4b, ord, &max);
+                       BUG_ON((start >> ord) >= max);
+                       mb_set_bit(start >> ord, buddy);
+                       e4b->bd_info->bb_counters[ord]--;
+                       start += mlen;
+                       len -= mlen;
+                       BUG_ON(len < 0);
+                       continue;
+               }
+
+               /* store for history */
+               if (ret == 0)
+                       ret = len | (ord << 16);
+
+               /* we have to split large buddy */
+               BUG_ON(ord <= 0);
+               buddy = mb_find_buddy(e4b, ord, &max);
+               mb_set_bit(start >> ord, buddy);
+               e4b->bd_info->bb_counters[ord]--;
+
+               ord--;
+               cur = (start >> ord) & ~1U;
+               buddy = mb_find_buddy(e4b, ord, &max);
+               mb_clear_bit(cur, buddy);
+               mb_clear_bit(cur + 1, buddy);
+               e4b->bd_info->bb_counters[ord]++;
+               e4b->bd_info->bb_counters[ord]++;
+       }
+
+       mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group),
+                       EXT4_MB_BITMAP(e4b), ex->fe_start, len0);
+       mb_check_buddy(e4b);
+
+       return ret;
+}
+
+/*
+ * Must be called under group lock!
+ */
+static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
+                                       struct ext4_buddy *e4b)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+       int ret;
+
+       BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
+       BUG_ON(ac->ac_status == AC_STATUS_FOUND);
+
+       ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
+       ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
+       ret = mb_mark_used(e4b, &ac->ac_b_ex);
+
+       /* preallocation can change ac_b_ex, thus we store actually
+        * allocated blocks for history */
+       ac->ac_f_ex = ac->ac_b_ex;
+
+       ac->ac_status = AC_STATUS_FOUND;
+       ac->ac_tail = ret & 0xffff;
+       ac->ac_buddy = ret >> 16;
+
+       /* XXXXXXX: SUCH A HORRIBLE **CK */
+       /*FIXME!! Why ? */
+       ac->ac_bitmap_page = e4b->bd_bitmap_page;
+       get_page(ac->ac_bitmap_page);
+       ac->ac_buddy_page = e4b->bd_buddy_page;
+       get_page(ac->ac_buddy_page);
+
+       /* store last allocated for subsequent stream allocation */
+       if ((ac->ac_flags & EXT4_MB_HINT_DATA)) {
+               spin_lock(&sbi->s_md_lock);
+               sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
+               sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
+               spin_unlock(&sbi->s_md_lock);
+       }
+}
+
+/*
+ * regular allocator, for general purposes allocation
+ */
+
+static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
+                                       struct ext4_buddy *e4b,
+                                       int finish_group)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+       struct ext4_free_extent *bex = &ac->ac_b_ex;
+       struct ext4_free_extent *gex = &ac->ac_g_ex;
+       struct ext4_free_extent ex;
+       int max;
+
+       /*
+        * We don't want to scan for a whole year
+        */
+       if (ac->ac_found > sbi->s_mb_max_to_scan &&
+                       !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
+               ac->ac_status = AC_STATUS_BREAK;
+               return;
+       }
+
+       /*
+        * Haven't found good chunk so far, let's continue
+        */
+       if (bex->fe_len < gex->fe_len)
+               return;
+
+       if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
+                       && bex->fe_group == e4b->bd_group) {
+               /* recheck chunk's availability - we don't know
+                * when it was found (within this lock-unlock
+                * period or not) */
+               max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
+               if (max >= gex->fe_len) {
+                       ext4_mb_use_best_found(ac, e4b);
+                       return;
+               }
+       }
+}
+
+/*
+ * The routine checks whether found extent is good enough. If it is,
+ * then the extent gets marked used and flag is set to the context
+ * to stop scanning. Otherwise, the extent is compared with the
+ * previous found extent and if new one is better, then it's stored
+ * in the context. Later, the best found extent will be used, if
+ * mballoc can't find good enough extent.
+ *
+ * FIXME: real allocation policy is to be designed yet!
+ */
+static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
+                                       struct ext4_free_extent *ex,
+                                       struct ext4_buddy *e4b)
+{
+       struct ext4_free_extent *bex = &ac->ac_b_ex;
+       struct ext4_free_extent *gex = &ac->ac_g_ex;
+
+       BUG_ON(ex->fe_len <= 0);
+       BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
+       BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
+       BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
+
+       ac->ac_found++;
+
+       /*
+        * The special case - take what you catch first
+        */
+       if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
+               *bex = *ex;
+               ext4_mb_use_best_found(ac, e4b);
+               return;
+       }
+
+       /*
+        * Let's check whether the chuck is good enough
+        */
+       if (ex->fe_len == gex->fe_len) {
+               *bex = *ex;
+               ext4_mb_use_best_found(ac, e4b);
+               return;
+       }
+
+       /*
+        * If this is first found extent, just store it in the context
+        */
+       if (bex->fe_len == 0) {
+               *bex = *ex;
+               return;
+       }
+
+       /*
+        * If new found extent is better, store it in the context
+        */
+       if (bex->fe_len < gex->fe_len) {
+               /* if the request isn't satisfied, any found extent
+                * larger than previous best one is better */
+               if (ex->fe_len > bex->fe_len)
+                       *bex = *ex;
+       } else if (ex->fe_len > gex->fe_len) {
+               /* if the request is satisfied, then we try to find
+                * an extent that still satisfy the request, but is
+                * smaller than previous one */
+               if (ex->fe_len < bex->fe_len)
+                       *bex = *ex;
+       }
+
+       ext4_mb_check_limits(ac, e4b, 0);
+}
+
+static int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
+                                       struct ext4_buddy *e4b)
+{
+       struct ext4_free_extent ex = ac->ac_b_ex;
+       ext4_group_t group = ex.fe_group;
+       int max;
+       int err;
+
+       BUG_ON(ex.fe_len <= 0);
+       err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
+       if (err)
+               return err;
+
+       ext4_lock_group(ac->ac_sb, group);
+       max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
+
+       if (max > 0) {
+               ac->ac_b_ex = ex;
+               ext4_mb_use_best_found(ac, e4b);
+       }
+
+       ext4_unlock_group(ac->ac_sb, group);
+       ext4_mb_release_desc(e4b);
+
+       return 0;
+}
+
+static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
+                               struct ext4_buddy *e4b)
+{
+       ext4_group_t group = ac->ac_g_ex.fe_group;
+       int max;
+       int err;
+       struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+       struct ext4_super_block *es = sbi->s_es;
+       struct ext4_free_extent ex;
+
+       if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
+               return 0;
+
+       err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
+       if (err)
+               return err;
+
+       ext4_lock_group(ac->ac_sb, group);
+       max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
+                            ac->ac_g_ex.fe_len, &ex);
+
+       if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
+               ext4_fsblk_t start;
+
+               start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) +
+                       ex.fe_start + le32_to_cpu(es->s_first_data_block);
+               /* use do_div to get remainder (would be 64-bit modulo) */
+               if (do_div(start, sbi->s_stripe) == 0) {
+                       ac->ac_found++;
+                       ac->ac_b_ex = ex;
+                       ext4_mb_use_best_found(ac, e4b);
+               }
+       } else if (max >= ac->ac_g_ex.fe_len) {
+               BUG_ON(ex.fe_len <= 0);
+               BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
+               BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
+               ac->ac_found++;
+               ac->ac_b_ex = ex;
+               ext4_mb_use_best_found(ac, e4b);
+       } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
+               /* Sometimes, caller may want to merge even small
+                * number of blocks to an existing extent */
+               BUG_ON(ex.fe_len <= 0);
+               BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
+               BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
+               ac->ac_found++;
+               ac->ac_b_ex = ex;
+               ext4_mb_use_best_found(ac, e4b);
+       }
+       ext4_unlock_group(ac->ac_sb, group);
+       ext4_mb_release_desc(e4b);
+
+       return 0;
+}
+
+/*
+ * The routine scans buddy structures (not bitmap!) from given order
+ * to max order and tries to find big enough chunk to satisfy the req
+ */
+static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
+                                       struct ext4_buddy *e4b)
+{
+       struct super_block *sb = ac->ac_sb;
+       struct ext4_group_info *grp = e4b->bd_info;
+       void *buddy;
+       int i;
+       int k;
+       int max;
+
+       BUG_ON(ac->ac_2order <= 0);
+       for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
+               if (grp->bb_counters[i] == 0)
+                       continue;
+
+               buddy = mb_find_buddy(e4b, i, &max);
+               BUG_ON(buddy == NULL);
+
+               k = ext4_find_next_zero_bit(buddy, max, 0);
+               BUG_ON(k >= max);
+
+               ac->ac_found++;
+
+               ac->ac_b_ex.fe_len = 1 << i;
+               ac->ac_b_ex.fe_start = k << i;
+               ac->ac_b_ex.fe_group = e4b->bd_group;
+
+               ext4_mb_use_best_found(ac, e4b);
+
+               BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
+
+               if (EXT4_SB(sb)->s_mb_stats)
+                       atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
+
+               break;
+       }
+}
+
+/*
+ * The routine scans the group and measures all found extents.
+ * In order to optimize scanning, caller must pass number of
+ * free blocks in the group, so the routine can know upper limit.
+ */
+static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
+                                       struct ext4_buddy *e4b)
+{
+       struct super_block *sb = ac->ac_sb;
+       void *bitmap = EXT4_MB_BITMAP(e4b);
+       struct ext4_free_extent ex;
+       int i;
+       int free;
+
+       free = e4b->bd_info->bb_free;
+       BUG_ON(free <= 0);
+
+       i = e4b->bd_info->bb_first_free;
+
+       while (free && ac->ac_status == AC_STATUS_CONTINUE) {
+               i = ext4_find_next_zero_bit(bitmap,
+                                               EXT4_BLOCKS_PER_GROUP(sb), i);
+               if (i >= EXT4_BLOCKS_PER_GROUP(sb)) {
+                       BUG_ON(free != 0);
+                       break;
+               }
+
+               mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
+               BUG_ON(ex.fe_len <= 0);
+               BUG_ON(free < ex.fe_len);
+
+               ext4_mb_measure_extent(ac, &ex, e4b);
+
+               i += ex.fe_len;
+               free -= ex.fe_len;
+       }
+
+       ext4_mb_check_limits(ac, e4b, 1);
+}
+
+/*
+ * This is a special case for storages like raid5
+ * we try to find stripe-aligned chunks for stripe-size requests
+ * XXX should do so at least for multiples of stripe size as well
+ */
+static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
+                                struct ext4_buddy *e4b)
+{
+       struct super_block *sb = ac->ac_sb;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       void *bitmap = EXT4_MB_BITMAP(e4b);
+       struct ext4_free_extent ex;
+       ext4_fsblk_t first_group_block;
+       ext4_fsblk_t a;
+       ext4_grpblk_t i;
+       int max;
+
+       BUG_ON(sbi->s_stripe == 0);
+
+       /* find first stripe-aligned block in group */
+       first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
+               + le32_to_cpu(sbi->s_es->s_first_data_block);
+       a = first_group_block + sbi->s_stripe - 1;
+       do_div(a, sbi->s_stripe);
+       i = (a * sbi->s_stripe) - first_group_block;
+
+       while (i < EXT4_BLOCKS_PER_GROUP(sb)) {
+               if (!mb_test_bit(i, bitmap)) {
+                       max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
+                       if (max >= sbi->s_stripe) {
+                               ac->ac_found++;
+                               ac->ac_b_ex = ex;
+                               ext4_mb_use_best_found(ac, e4b);
+                               break;
+                       }
+               }
+               i += sbi->s_stripe;
+       }
+}
+
+static int ext4_mb_good_group(struct ext4_allocation_context *ac,
+                               ext4_group_t group, int cr)
+{
+       unsigned free, fragments;
+       unsigned i, bits;
+       struct ext4_group_desc *desc;
+       struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
+
+       BUG_ON(cr < 0 || cr >= 4);
+       BUG_ON(EXT4_MB_GRP_NEED_INIT(grp));
+
+       free = grp->bb_free;
+       fragments = grp->bb_fragments;
+       if (free == 0)
+               return 0;
+       if (fragments == 0)
+               return 0;
+
+       switch (cr) {
+       case 0:
+               BUG_ON(ac->ac_2order == 0);
+               /* If this group is uninitialized, skip it initially */
+               desc = ext4_get_group_desc(ac->ac_sb, group, NULL);
+               if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
+                       return 0;
+
+               bits = ac->ac_sb->s_blocksize_bits + 1;
+               for (i = ac->ac_2order; i <= bits; i++)
+                       if (grp->bb_counters[i] > 0)
+                               return 1;
+               break;
+       case 1:
+               if ((free / fragments) >= ac->ac_g_ex.fe_len)
+                       return 1;
+               break;
+       case 2:
+               if (free >= ac->ac_g_ex.fe_len)
+                       return 1;
+               break;
+       case 3:
+               return 1;
+       default:
+               BUG();
+       }
+
+       return 0;
+}
+
+static int ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
+{
+       ext4_group_t group;
+       ext4_group_t i;
+       int cr;
+       int err = 0;
+       int bsbits;
+       struct ext4_sb_info *sbi;
+       struct super_block *sb;
+       struct ext4_buddy e4b;
+       loff_t size, isize;
+
+       sb = ac->ac_sb;
+       sbi = EXT4_SB(sb);
+       BUG_ON(ac->ac_status == AC_STATUS_FOUND);
+
+       /* first, try the goal */
+       err = ext4_mb_find_by_goal(ac, &e4b);
+       if (err || ac->ac_status == AC_STATUS_FOUND)
+               goto out;
+
+       if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
+               goto out;
+
+       /*
+        * ac->ac2_order is set only if the fe_len is a power of 2
+        * if ac2_order is set we also set criteria to 0 so that we
+        * try exact allocation using buddy.
+        */
+       i = fls(ac->ac_g_ex.fe_len);
+       ac->ac_2order = 0;
+       /*
+        * We search using buddy data only if the order of the request
+        * is greater than equal to the sbi_s_mb_order2_reqs
+        * You can tune it via /proc/fs/ext4/<partition>/order2_req
+        */
+       if (i >= sbi->s_mb_order2_reqs) {
+               /*
+                * This should tell if fe_len is exactly power of 2
+                */
+               if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
+                       ac->ac_2order = i - 1;
+       }
+
+       bsbits = ac->ac_sb->s_blocksize_bits;
+       /* if stream allocation is enabled, use global goal */
+       size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
+       isize = i_size_read(ac->ac_inode) >> bsbits;
+       if (size < isize)
+               size = isize;
+
+       if (size < sbi->s_mb_stream_request &&
+                       (ac->ac_flags & EXT4_MB_HINT_DATA)) {
+               /* TBD: may be hot point */
+               spin_lock(&sbi->s_md_lock);
+               ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
+               ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
+               spin_unlock(&sbi->s_md_lock);
+       }
+
+       /* searching for the right group start from the goal value specified */
+       group = ac->ac_g_ex.fe_group;
+
+       /* Let's just scan groups to find more-less suitable blocks */
+       cr = ac->ac_2order ? 0 : 1;
+       /*
+        * cr == 0 try to get exact allocation,
+        * cr == 3  try to get anything
+        */
+repeat:
+       for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
+               ac->ac_criteria = cr;
+               for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) {
+                       struct ext4_group_info *grp;
+                       struct ext4_group_desc *desc;
+
+                       if (group == EXT4_SB(sb)->s_groups_count)
+                               group = 0;
+
+                       /* quick check to skip empty groups */
+                       grp = ext4_get_group_info(ac->ac_sb, group);
+                       if (grp->bb_free == 0)
+                               continue;
+
+                       /*
+                        * if the group is already init we check whether it is
+                        * a good group and if not we don't load the buddy
+                        */
+                       if (EXT4_MB_GRP_NEED_INIT(grp)) {
+                               /*
+                                * we need full data about the group
+                                * to make a good selection
+                                */
+                               err = ext4_mb_load_buddy(sb, group, &e4b);
+                               if (err)
+                                       goto out;
+                               ext4_mb_release_desc(&e4b);
+                       }
+
+                       /*
+                        * If the particular group doesn't satisfy our
+                        * criteria we continue with the next group
+                        */
+                       if (!ext4_mb_good_group(ac, group, cr))
+                               continue;
+
+                       err = ext4_mb_load_buddy(sb, group, &e4b);
+                       if (err)
+                               goto out;
+
+                       ext4_lock_group(sb, group);
+                       if (!ext4_mb_good_group(ac, group, cr)) {
+                               /* someone did allocation from this group */
+                               ext4_unlock_group(sb, group);
+                               ext4_mb_release_desc(&e4b);
+                               continue;
+                       }
+
+                       ac->ac_groups_scanned++;
+                       desc = ext4_get_group_desc(sb, group, NULL);
+                       if (cr == 0 || (desc->bg_flags &
+                                       cpu_to_le16(EXT4_BG_BLOCK_UNINIT) &&
+                                       ac->ac_2order != 0))
+                               ext4_mb_simple_scan_group(ac, &e4b);
+                       else if (cr == 1 &&
+                                       ac->ac_g_ex.fe_len == sbi->s_stripe)
+                               ext4_mb_scan_aligned(ac, &e4b);
+                       else
+                               ext4_mb_complex_scan_group(ac, &e4b);
+
+                       ext4_unlock_group(sb, group);
+                       ext4_mb_release_desc(&e4b);
+
+                       if (ac->ac_status != AC_STATUS_CONTINUE)
+                               break;
+               }
+       }
+
+       if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
+           !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
+               /*
+                * We've been searching too long. Let's try to allocate
+                * the best chunk we've found so far
+                */
+
+               ext4_mb_try_best_found(ac, &e4b);
+               if (ac->ac_status != AC_STATUS_FOUND) {
+                       /*
+                        * Someone more lucky has already allocated it.
+                        * The only thing we can do is just take first
+                        * found block(s)
+                       printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
+                        */
+                       ac->ac_b_ex.fe_group = 0;
+                       ac->ac_b_ex.fe_start = 0;
+                       ac->ac_b_ex.fe_len = 0;
+                       ac->ac_status = AC_STATUS_CONTINUE;
+                       ac->ac_flags |= EXT4_MB_HINT_FIRST;
+                       cr = 3;
+                       atomic_inc(&sbi->s_mb_lost_chunks);
+                       goto repeat;
+               }
+       }
+out:
+       return err;
+}
+
+#ifdef EXT4_MB_HISTORY
+struct ext4_mb_proc_session {
+       struct ext4_mb_history *history;
+       struct super_block *sb;
+       int start;
+       int max;
+};
+
+static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s,
+                                       struct ext4_mb_history *hs,
+                                       int first)
+{
+       if (hs == s->history + s->max)
+               hs = s->history;
+       if (!first && hs == s->history + s->start)
+               return NULL;
+       while (hs->orig.fe_len == 0) {
+               hs++;
+               if (hs == s->history + s->max)
+                       hs = s->history;
+               if (hs == s->history + s->start)
+                       return NULL;
+       }
+       return hs;
+}
+
+static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos)
+{
+       struct ext4_mb_proc_session *s = seq->private;
+       struct ext4_mb_history *hs;
+       int l = *pos;
+
+       if (l == 0)
+               return SEQ_START_TOKEN;
+       hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1);
+       if (!hs)
+               return NULL;
+       while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL);
+       return hs;
+}
+
+static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v,
+                                     loff_t *pos)
+{
+       struct ext4_mb_proc_session *s = seq->private;
+       struct ext4_mb_history *hs = v;
+
+       ++*pos;
+       if (v == SEQ_START_TOKEN)
+               return ext4_mb_history_skip_empty(s, s->history + s->start, 1);
+       else
+               return ext4_mb_history_skip_empty(s, ++hs, 0);
+}
+
+static int ext4_mb_seq_history_show(struct seq_file *seq, void *v)
+{
+       char buf[25], buf2[25], buf3[25], *fmt;
+       struct ext4_mb_history *hs = v;
+
+       if (v == SEQ_START_TOKEN) {
+               seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s "
+                               "%-5s %-2s %-5s %-5s %-5s %-6s\n",
+                         "pid", "inode", "original", "goal", "result", "found",
+                          "grps", "cr", "flags", "merge", "tail", "broken");
+               return 0;
+       }
+
+       if (hs->op == EXT4_MB_HISTORY_ALLOC) {
+               fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u "
+                       "%-5u %-5s %-5u %-6u\n";
+               sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
+                       hs->result.fe_start, hs->result.fe_len,
+                       hs->result.fe_logical);
+               sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
+                       hs->orig.fe_start, hs->orig.fe_len,
+                       hs->orig.fe_logical);
+               sprintf(buf3, "%lu/%d/%u@%u", hs->goal.fe_group,
+                       hs->goal.fe_start, hs->goal.fe_len,
+                       hs->goal.fe_logical);
+               seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2,
+                               hs->found, hs->groups, hs->cr, hs->flags,
+                               hs->merged ? "M" : "", hs->tail,
+                               hs->buddy ? 1 << hs->buddy : 0);
+       } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) {
+               fmt = "%-5u %-8u %-23s %-23s %-23s\n";
+               sprintf(buf2, "%lu/%d/%u@%u", hs->result.fe_group,
+                       hs->result.fe_start, hs->result.fe_len,
+                       hs->result.fe_logical);
+               sprintf(buf, "%lu/%d/%u@%u", hs->orig.fe_group,
+                       hs->orig.fe_start, hs->orig.fe_len,
+                       hs->orig.fe_logical);
+               seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2);
+       } else if (hs->op == EXT4_MB_HISTORY_DISCARD) {
+               sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
+                       hs->result.fe_start, hs->result.fe_len);
+               seq_printf(seq, "%-5u %-8u %-23s discard\n",
+                               hs->pid, hs->ino, buf2);
+       } else if (hs->op == EXT4_MB_HISTORY_FREE) {
+               sprintf(buf2, "%lu/%d/%u", hs->result.fe_group,
+                       hs->result.fe_start, hs->result.fe_len);
+               seq_printf(seq, "%-5u %-8u %-23s free\n",
+                               hs->pid, hs->ino, buf2);
+       }
+       return 0;
+}
+
+static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v)
+{
+}
+
+static struct seq_operations ext4_mb_seq_history_ops = {
+       .start  = ext4_mb_seq_history_start,
+       .next   = ext4_mb_seq_history_next,
+       .stop   = ext4_mb_seq_history_stop,
+       .show   = ext4_mb_seq_history_show,
+};
+
+static int ext4_mb_seq_history_open(struct inode *inode, struct file *file)
+{
+       struct super_block *sb = PDE(inode)->data;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       struct ext4_mb_proc_session *s;
+       int rc;
+       int size;
+
+       s = kmalloc(sizeof(*s), GFP_KERNEL);
+       if (s == NULL)
+               return -ENOMEM;
+       s->sb = sb;
+       size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max;
+       s->history = kmalloc(size, GFP_KERNEL);
+       if (s->history == NULL) {
+               kfree(s);
+               return -ENOMEM;
+       }
+
+       spin_lock(&sbi->s_mb_history_lock);
+       memcpy(s->history, sbi->s_mb_history, size);
+       s->max = sbi->s_mb_history_max;
+       s->start = sbi->s_mb_history_cur % s->max;
+       spin_unlock(&sbi->s_mb_history_lock);
+
+       rc = seq_open(file, &ext4_mb_seq_history_ops);
+       if (rc == 0) {
+               struct seq_file *m = (struct seq_file *)file->private_data;
+               m->private = s;
+       } else {
+               kfree(s->history);
+               kfree(s);
+       }
+       return rc;
+
+}
+
+static int ext4_mb_seq_history_release(struct inode *inode, struct file *file)
+{
+       struct seq_file *seq = (struct seq_file *)file->private_data;
+       struct ext4_mb_proc_session *s = seq->private;
+       kfree(s->history);
+       kfree(s);
+       return seq_release(inode, file);
+}
+
+static ssize_t ext4_mb_seq_history_write(struct file *file,
+                               const char __user *buffer,
+                               size_t count, loff_t *ppos)
+{
+       struct seq_file *seq = (struct seq_file *)file->private_data;
+       struct ext4_mb_proc_session *s = seq->private;
+       struct super_block *sb = s->sb;
+       char str[32];
+       int value;
+
+       if (count >= sizeof(str)) {
+               printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n",
+                               "mb_history", (int)sizeof(str));
+               return -EOVERFLOW;
+       }
+
+       if (copy_from_user(str, buffer, count))
+               return -EFAULT;
+
+       value = simple_strtol(str, NULL, 0);
+       if (value < 0)
+               return -ERANGE;
+       EXT4_SB(sb)->s_mb_history_filter = value;
+
+       return count;
+}
+
+static struct file_operations ext4_mb_seq_history_fops = {
+       .owner          = THIS_MODULE,
+       .open           = ext4_mb_seq_history_open,
+       .read           = seq_read,
+       .write          = ext4_mb_seq_history_write,
+       .llseek         = seq_lseek,
+       .release        = ext4_mb_seq_history_release,
+};
+
+static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
+{
+       struct super_block *sb = seq->private;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       ext4_group_t group;
+
+       if (*pos < 0 || *pos >= sbi->s_groups_count)
+               return NULL;
+
+       group = *pos + 1;
+       return (void *) group;
+}
+
+static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
+{
+       struct super_block *sb = seq->private;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       ext4_group_t group;
+
+       ++*pos;
+       if (*pos < 0 || *pos >= sbi->s_groups_count)
+               return NULL;
+       group = *pos + 1;
+       return (void *) group;;
+}
+
+static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
+{
+       struct super_block *sb = seq->private;
+       long group = (long) v;
+       int i;
+       int err;
+       struct ext4_buddy e4b;
+       struct sg {
+               struct ext4_group_info info;
+               unsigned short counters[16];
+       } sg;
+
+       group--;
+       if (group == 0)
+               seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
+                               "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
+                                 "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
+                          "group", "free", "frags", "first",
+                          "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
+                          "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
+
+       i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
+               sizeof(struct ext4_group_info);
+       err = ext4_mb_load_buddy(sb, group, &e4b);
+       if (err) {
+               seq_printf(seq, "#%-5lu: I/O error\n", group);
+               return 0;
+       }
+       ext4_lock_group(sb, group);
+       memcpy(&sg, ext4_get_group_info(sb, group), i);
+       ext4_unlock_group(sb, group);
+       ext4_mb_release_desc(&e4b);
+
+       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
+                       sg.info.bb_fragments, sg.info.bb_first_free);
+       for (i = 0; i <= 13; i++)
+               seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
+                               sg.info.bb_counters[i] : 0);
+       seq_printf(seq, " ]\n");
+
+       return 0;
+}
+
+static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
+{
+}
+
+static struct seq_operations ext4_mb_seq_groups_ops = {
+       .start  = ext4_mb_seq_groups_start,
+       .next   = ext4_mb_seq_groups_next,
+       .stop   = ext4_mb_seq_groups_stop,
+       .show   = ext4_mb_seq_groups_show,
+};
+
+static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
+{
+       struct super_block *sb = PDE(inode)->data;
+       int rc;
+
+       rc = seq_open(file, &ext4_mb_seq_groups_ops);
+       if (rc == 0) {
+               struct seq_file *m = (struct seq_file *)file->private_data;
+               m->private = sb;
+       }
+       return rc;
+
+}
+
+static struct file_operations ext4_mb_seq_groups_fops = {
+       .owner          = THIS_MODULE,
+       .open           = ext4_mb_seq_groups_open,
+       .read           = seq_read,
+       .llseek         = seq_lseek,
+       .release        = seq_release,
+};
+
+static void ext4_mb_history_release(struct super_block *sb)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+
+       remove_proc_entry("mb_groups", sbi->s_mb_proc);
+       remove_proc_entry("mb_history", sbi->s_mb_proc);
+
+       kfree(sbi->s_mb_history);
+}
+
+static void ext4_mb_history_init(struct super_block *sb)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       int i;
+
+       if (sbi->s_mb_proc != NULL) {
+               struct proc_dir_entry *p;
+               p = create_proc_entry("mb_history", S_IRUGO, sbi->s_mb_proc);
+               if (p) {
+                       p->proc_fops = &ext4_mb_seq_history_fops;
+                       p->data = sb;
+               }
+               p = create_proc_entry("mb_groups", S_IRUGO, sbi->s_mb_proc);
+               if (p) {
+                       p->proc_fops = &ext4_mb_seq_groups_fops;
+                       p->data = sb;
+               }
+       }
+
+       sbi->s_mb_history_max = 1000;
+       sbi->s_mb_history_cur = 0;
+       spin_lock_init(&sbi->s_mb_history_lock);
+       i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
+       sbi->s_mb_history = kmalloc(i, GFP_KERNEL);
+       if (likely(sbi->s_mb_history != NULL))
+               memset(sbi->s_mb_history, 0, i);
+       /* if we can't allocate history, then we simple won't use it */
+}
+
+static void ext4_mb_store_history(struct ext4_allocation_context *ac)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+       struct ext4_mb_history h;
+
+       if (unlikely(sbi->s_mb_history == NULL))
+               return;
+
+       if (!(ac->ac_op & sbi->s_mb_history_filter))
+               return;
+
+       h.op = ac->ac_op;
+       h.pid = current->pid;
+       h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0;
+       h.orig = ac->ac_o_ex;
+       h.result = ac->ac_b_ex;
+       h.flags = ac->ac_flags;
+       h.found = ac->ac_found;
+       h.groups = ac->ac_groups_scanned;
+       h.cr = ac->ac_criteria;
+       h.tail = ac->ac_tail;
+       h.buddy = ac->ac_buddy;
+       h.merged = 0;
+       if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
+               if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
+                               ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
+                       h.merged = 1;
+               h.goal = ac->ac_g_ex;
+               h.result = ac->ac_f_ex;
+       }
+
+       spin_lock(&sbi->s_mb_history_lock);
+       memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h));
+       if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max)
+               sbi->s_mb_history_cur = 0;
+       spin_unlock(&sbi->s_mb_history_lock);
+}
+
+#else
+#define ext4_mb_history_release(sb)
+#define ext4_mb_history_init(sb)
+#endif
+
+static int ext4_mb_init_backend(struct super_block *sb)
+{
+       ext4_group_t i;
+       int j, len, metalen;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       int num_meta_group_infos =
+               (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) - 1) >>
+                       EXT4_DESC_PER_BLOCK_BITS(sb);
+       struct ext4_group_info **meta_group_info;
+
+       /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
+        * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
+        * So a two level scheme suffices for now. */
+       sbi->s_group_info = kmalloc(sizeof(*sbi->s_group_info) *
+                                   num_meta_group_infos, GFP_KERNEL);
+       if (sbi->s_group_info == NULL) {
+               printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n");
+               return -ENOMEM;
+       }
+       sbi->s_buddy_cache = new_inode(sb);
+       if (sbi->s_buddy_cache == NULL) {
+               printk(KERN_ERR "EXT4-fs: can't get new inode\n");
+               goto err_freesgi;
+       }
+       EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
+
+       metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb);
+       for (i = 0; i < num_meta_group_infos; i++) {
+               if ((i + 1) == num_meta_group_infos)
+                       metalen = sizeof(*meta_group_info) *
+                               (sbi->s_groups_count -
+                                       (i << EXT4_DESC_PER_BLOCK_BITS(sb)));
+               meta_group_info = kmalloc(metalen, GFP_KERNEL);
+               if (meta_group_info == NULL) {
+                       printk(KERN_ERR "EXT4-fs: can't allocate mem for a "
+                              "buddy group\n");
+                       goto err_freemeta;
+               }
+               sbi->s_group_info[i] = meta_group_info;
+       }
+
+       /*
+        * calculate needed size. if change bb_counters size,
+        * don't forget about ext4_mb_generate_buddy()
+        */
+       len = sizeof(struct ext4_group_info);
+       len += sizeof(unsigned short) * (sb->s_blocksize_bits + 2);
+       for (i = 0; i < sbi->s_groups_count; i++) {
+               struct ext4_group_desc *desc;
+
+               meta_group_info =
+                       sbi->s_group_info[i >> EXT4_DESC_PER_BLOCK_BITS(sb)];
+               j = i & (EXT4_DESC_PER_BLOCK(sb) - 1);
+
+               meta_group_info[j] = kzalloc(len, GFP_KERNEL);
+               if (meta_group_info[j] == NULL) {
+                       printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n");
+                       i--;
+                       goto err_freebuddy;
+               }
+               desc = ext4_get_group_desc(sb, i, NULL);
+               if (desc == NULL) {
+                       printk(KERN_ERR
+                               "EXT4-fs: can't read descriptor %lu\n", i);
+                       goto err_freebuddy;
+               }
+               memset(meta_group_info[j], 0, len);
+               set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
+                       &(meta_group_info[j]->bb_state));
+
+               /*
+                * initialize bb_free to be able to skip
+                * empty groups without initialization
+                */
+               if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
+                       meta_group_info[j]->bb_free =
+                               ext4_free_blocks_after_init(sb, i, desc);
+               } else {
+                       meta_group_info[j]->bb_free =
+                               le16_to_cpu(desc->bg_free_blocks_count);
+               }
+
+               INIT_LIST_HEAD(&meta_group_info[j]->bb_prealloc_list);
+
+#ifdef DOUBLE_CHECK
+               {
+                       struct buffer_head *bh;
+                       meta_group_info[j]->bb_bitmap =
+                               kmalloc(sb->s_blocksize, GFP_KERNEL);
+                       BUG_ON(meta_group_info[j]->bb_bitmap == NULL);
+                       bh = read_block_bitmap(sb, i);
+                       BUG_ON(bh == NULL);
+                       memcpy(meta_group_info[j]->bb_bitmap, bh->b_data,
+                                       sb->s_blocksize);
+                       put_bh(bh);
+               }
+#endif
+
+       }
+
+       return 0;
+
+err_freebuddy:
+       while (i >= 0) {
+               kfree(ext4_get_group_info(sb, i));
+               i--;
+       }
+       i = num_meta_group_infos;
+err_freemeta:
+       while (--i >= 0)
+               kfree(sbi->s_group_info[i]);
+       iput(sbi->s_buddy_cache);
+err_freesgi:
+       kfree(sbi->s_group_info);
+       return -ENOMEM;
+}
+
+int ext4_mb_init(struct super_block *sb, int needs_recovery)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       unsigned i;
+       unsigned offset;
+       unsigned max;
+
+       if (!test_opt(sb, MBALLOC))
+               return 0;
+
+       i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short);
+
+       sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
+       if (sbi->s_mb_offsets == NULL) {
+               clear_opt(sbi->s_mount_opt, MBALLOC);
+               return -ENOMEM;
+       }
+       sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
+       if (sbi->s_mb_maxs == NULL) {
+               clear_opt(sbi->s_mount_opt, MBALLOC);
+               kfree(sbi->s_mb_maxs);
+               return -ENOMEM;
+       }
+
+       /* order 0 is regular bitmap */
+       sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
+       sbi->s_mb_offsets[0] = 0;
+
+       i = 1;
+       offset = 0;
+       max = sb->s_blocksize << 2;
+       do {
+               sbi->s_mb_offsets[i] = offset;
+               sbi->s_mb_maxs[i] = max;
+               offset += 1 << (sb->s_blocksize_bits - i);
+               max = max >> 1;
+               i++;
+       } while (i <= sb->s_blocksize_bits + 1);
+
+       /* init file for buddy data */
+       i = ext4_mb_init_backend(sb);
+       if (i) {
+               clear_opt(sbi->s_mount_opt, MBALLOC);
+               kfree(sbi->s_mb_offsets);
+               kfree(sbi->s_mb_maxs);
+               return i;
+       }
+
+       spin_lock_init(&sbi->s_md_lock);
+       INIT_LIST_HEAD(&sbi->s_active_transaction);
+       INIT_LIST_HEAD(&sbi->s_closed_transaction);
+       INIT_LIST_HEAD(&sbi->s_committed_transaction);
+       spin_lock_init(&sbi->s_bal_lock);
+
+       sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
+       sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
+       sbi->s_mb_stats = MB_DEFAULT_STATS;
+       sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
+       sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
+       sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT;
+       sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
+
+       i = sizeof(struct ext4_locality_group) * NR_CPUS;
+       sbi->s_locality_groups = kmalloc(i, GFP_KERNEL);
+       if (sbi->s_locality_groups == NULL) {
+               clear_opt(sbi->s_mount_opt, MBALLOC);
+               kfree(sbi->s_mb_offsets);
+               kfree(sbi->s_mb_maxs);
+               return -ENOMEM;
+       }
+       for (i = 0; i < NR_CPUS; i++) {
+               struct ext4_locality_group *lg;
+               lg = &sbi->s_locality_groups[i];
+               mutex_init(&lg->lg_mutex);
+               INIT_LIST_HEAD(&lg->lg_prealloc_list);
+               spin_lock_init(&lg->lg_prealloc_lock);
+       }
+
+       ext4_mb_init_per_dev_proc(sb);
+       ext4_mb_history_init(sb);
+
+       printk("EXT4-fs: mballoc enabled\n");
+       return 0;
+}
+
+/* need to called with ext4 group lock (ext4_lock_group) */
+static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
+{
+       struct ext4_prealloc_space *pa;
+       struct list_head *cur, *tmp;
+       int count = 0;
+
+       list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
+               pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
+               list_del(&pa->pa_group_list);
+               count++;
+               kfree(pa);
+       }
+       if (count)
+               mb_debug("mballoc: %u PAs left\n", count);
+
+}
+
+int ext4_mb_release(struct super_block *sb)
+{
+       ext4_group_t i;
+       int num_meta_group_infos;
+       struct ext4_group_info *grinfo;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+
+       if (!test_opt(sb, MBALLOC))
+               return 0;
+
+       /* release freed, non-committed blocks */
+       spin_lock(&sbi->s_md_lock);
+       list_splice_init(&sbi->s_closed_transaction,
+                       &sbi->s_committed_transaction);
+       list_splice_init(&sbi->s_active_transaction,
+                       &sbi->s_committed_transaction);
+       spin_unlock(&sbi->s_md_lock);
+       ext4_mb_free_committed_blocks(sb);
+
+       if (sbi->s_group_info) {
+               for (i = 0; i < sbi->s_groups_count; i++) {
+                       grinfo = ext4_get_group_info(sb, i);
+#ifdef DOUBLE_CHECK
+                       kfree(grinfo->bb_bitmap);
+#endif
+                       ext4_lock_group(sb, i);
+                       ext4_mb_cleanup_pa(grinfo);
+                       ext4_unlock_group(sb, i);
+                       kfree(grinfo);
+               }
+               num_meta_group_infos = (sbi->s_groups_count +
+                               EXT4_DESC_PER_BLOCK(sb) - 1) >>
+                       EXT4_DESC_PER_BLOCK_BITS(sb);
+               for (i = 0; i < num_meta_group_infos; i++)
+                       kfree(sbi->s_group_info[i]);
+               kfree(sbi->s_group_info);
+       }
+       kfree(sbi->s_mb_offsets);
+       kfree(sbi->s_mb_maxs);
+       if (sbi->s_buddy_cache)
+               iput(sbi->s_buddy_cache);
+       if (sbi->s_mb_stats) {
+               printk(KERN_INFO
+                      "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n",
+                               atomic_read(&sbi->s_bal_allocated),
+                               atomic_read(&sbi->s_bal_reqs),
+                               atomic_read(&sbi->s_bal_success));
+               printk(KERN_INFO
+                     "EXT4-fs: mballoc: %u extents scanned, %u goal hits, "
+                               "%u 2^N hits, %u breaks, %u lost\n",
+                               atomic_read(&sbi->s_bal_ex_scanned),
+                               atomic_read(&sbi->s_bal_goals),
+                               atomic_read(&sbi->s_bal_2orders),
+                               atomic_read(&sbi->s_bal_breaks),
+                               atomic_read(&sbi->s_mb_lost_chunks));
+               printk(KERN_INFO
+                      "EXT4-fs: mballoc: %lu generated and it took %Lu\n",
+                               sbi->s_mb_buddies_generated++,
+                               sbi->s_mb_generation_time);
+               printk(KERN_INFO
+                      "EXT4-fs: mballoc: %u preallocated, %u discarded\n",
+                               atomic_read(&sbi->s_mb_preallocated),
+                               atomic_read(&sbi->s_mb_discarded));
+       }
+
+       kfree(sbi->s_locality_groups);
+
+       ext4_mb_history_release(sb);
+       ext4_mb_destroy_per_dev_proc(sb);
+
+       return 0;
+}
+
+static void ext4_mb_free_committed_blocks(struct super_block *sb)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       int err;
+       int i;
+       int count = 0;
+       int count2 = 0;
+       struct ext4_free_metadata *md;
+       struct ext4_buddy e4b;
+
+       if (list_empty(&sbi->s_committed_transaction))
+               return;
+
+       /* there is committed blocks to be freed yet */
+       do {
+               /* get next array of blocks */
+               md = NULL;
+               spin_lock(&sbi->s_md_lock);
+               if (!list_empty(&sbi->s_committed_transaction)) {
+                       md = list_entry(sbi->s_committed_transaction.next,
+                                       struct ext4_free_metadata, list);
+                       list_del(&md->list);
+               }
+               spin_unlock(&sbi->s_md_lock);
+
+               if (md == NULL)
+                       break;
+
+               mb_debug("gonna free %u blocks in group %lu (0x%p):",
+                               md->num, md->group, md);
+
+               err = ext4_mb_load_buddy(sb, md->group, &e4b);
+               /* we expect to find existing buddy because it's pinned */
+               BUG_ON(err != 0);
+
+               /* there are blocks to put in buddy to make them really free */
+               count += md->num;
+               count2++;
+               ext4_lock_group(sb, md->group);
+               for (i = 0; i < md->num; i++) {
+                       mb_debug(" %u", md->blocks[i]);
+                       err = mb_free_blocks(NULL, &e4b, md->blocks[i], 1);
+                       BUG_ON(err != 0);
+               }
+               mb_debug("\n");
+               ext4_unlock_group(sb, md->group);
+
+               /* balance refcounts from ext4_mb_free_metadata() */
+               page_cache_release(e4b.bd_buddy_page);
+               page_cache_release(e4b.bd_bitmap_page);
+
+               kfree(md);
+               ext4_mb_release_desc(&e4b);
+
+       } while (md);
+
+       mb_debug("freed %u blocks in %u structures\n", count, count2);
+}
+
+#define EXT4_ROOT                      "ext4"
+#define EXT4_MB_STATS_NAME             "stats"
+#define EXT4_MB_MAX_TO_SCAN_NAME       "max_to_scan"
+#define EXT4_MB_MIN_TO_SCAN_NAME       "min_to_scan"
+#define EXT4_MB_ORDER2_REQ             "order2_req"
+#define EXT4_MB_STREAM_REQ             "stream_req"
+#define EXT4_MB_GROUP_PREALLOC         "group_prealloc"
+
+
+
+#define MB_PROC_VALUE_READ(name)                               \
+static int ext4_mb_read_##name(char *page, char **start,       \
+               off_t off, int count, int *eof, void *data)     \
+{                                                              \
+       struct ext4_sb_info *sbi = data;                        \
+       int len;                                                \
+       *eof = 1;                                               \
+       if (off != 0)                                           \
+               return 0;                                       \
+       len = sprintf(page, "%ld\n", sbi->s_mb_##name);         \
+       *start = page;                                          \
+       return len;                                             \
+}
+
+#define MB_PROC_VALUE_WRITE(name)                              \
+static int ext4_mb_write_##name(struct file *file,             \
+               const char __user *buf, unsigned long cnt, void *data)  \
+{                                                              \
+       struct ext4_sb_info *sbi = data;                        \
+       char str[32];                                           \
+       long value;                                             \
+       if (cnt >= sizeof(str))                                 \
+               return -EINVAL;                                 \
+       if (copy_from_user(str, buf, cnt))                      \
+               return -EFAULT;                                 \
+       value = simple_strtol(str, NULL, 0);                    \
+       if (value <= 0)                                         \
+               return -ERANGE;                                 \
+       sbi->s_mb_##name = value;                               \
+       return cnt;                                             \
+}
+
+MB_PROC_VALUE_READ(stats);
+MB_PROC_VALUE_WRITE(stats);
+MB_PROC_VALUE_READ(max_to_scan);
+MB_PROC_VALUE_WRITE(max_to_scan);
+MB_PROC_VALUE_READ(min_to_scan);
+MB_PROC_VALUE_WRITE(min_to_scan);
+MB_PROC_VALUE_READ(order2_reqs);
+MB_PROC_VALUE_WRITE(order2_reqs);
+MB_PROC_VALUE_READ(stream_request);
+MB_PROC_VALUE_WRITE(stream_request);
+MB_PROC_VALUE_READ(group_prealloc);
+MB_PROC_VALUE_WRITE(group_prealloc);
+
+#define        MB_PROC_HANDLER(name, var)                                      \
+do {                                                                   \
+       proc = create_proc_entry(name, mode, sbi->s_mb_proc);           \
+       if (proc == NULL) {                                             \
+               printk(KERN_ERR "EXT4-fs: can't to create %s\n", name); \
+               goto err_out;                                           \
+       }                                                               \
+       proc->data = sbi;                                               \
+       proc->read_proc  = ext4_mb_read_##var ;                         \
+       proc->write_proc = ext4_mb_write_##var;                         \
+} while (0)
+
+static int ext4_mb_init_per_dev_proc(struct super_block *sb)
+{
+       mode_t mode = S_IFREG | S_IRUGO | S_IWUSR;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       struct proc_dir_entry *proc;
+       char devname[64];
+
+       snprintf(devname, sizeof(devname) - 1, "%s",
+               bdevname(sb->s_bdev, devname));
+       sbi->s_mb_proc = proc_mkdir(devname, proc_root_ext4);
+
+       MB_PROC_HANDLER(EXT4_MB_STATS_NAME, stats);
+       MB_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, max_to_scan);
+       MB_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, min_to_scan);
+       MB_PROC_HANDLER(EXT4_MB_ORDER2_REQ, order2_reqs);
+       MB_PROC_HANDLER(EXT4_MB_STREAM_REQ, stream_request);
+       MB_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, group_prealloc);
+
+       return 0;
+
+err_out:
+       printk(KERN_ERR "EXT4-fs: Unable to create %s\n", devname);
+       remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
+       remove_proc_entry(devname, proc_root_ext4);
+       sbi->s_mb_proc = NULL;
+
+       return -ENOMEM;
+}
+
+static int ext4_mb_destroy_per_dev_proc(struct super_block *sb)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       char devname[64];
+
+       if (sbi->s_mb_proc == NULL)
+               return -EINVAL;
+
+       snprintf(devname, sizeof(devname) - 1, "%s",
+               bdevname(sb->s_bdev, devname));
+       remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_mb_proc);
+       remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_mb_proc);
+       remove_proc_entry(devname, proc_root_ext4);
+
+       return 0;
+}
+
+int __init init_ext4_mballoc(void)
+{
+       ext4_pspace_cachep =
+               kmem_cache_create("ext4_prealloc_space",
+                                    sizeof(struct ext4_prealloc_space),
+                                    0, SLAB_RECLAIM_ACCOUNT, NULL);
+       if (ext4_pspace_cachep == NULL)
+               return -ENOMEM;
+
+#ifdef CONFIG_PROC_FS
+       proc_root_ext4 = proc_mkdir(EXT4_ROOT, proc_root_fs);
+       if (proc_root_ext4 == NULL)
+               printk(KERN_ERR "EXT4-fs: Unable to create %s\n", EXT4_ROOT);
+#endif
+
+       return 0;
+}
+
+void exit_ext4_mballoc(void)
+{
+       /* XXX: synchronize_rcu(); */
+       kmem_cache_destroy(ext4_pspace_cachep);
+#ifdef CONFIG_PROC_FS
+       remove_proc_entry(EXT4_ROOT, proc_root_fs);
+#endif
+}
+
+
+/*
+ * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps
+ * Returns 0 if success or error code
+ */
+static int ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
+                               handle_t *handle)
+{
+       struct buffer_head *bitmap_bh = NULL;
+       struct ext4_super_block *es;
+       struct ext4_group_desc *gdp;
+       struct buffer_head *gdp_bh;
+       struct ext4_sb_info *sbi;
+       struct super_block *sb;
+       ext4_fsblk_t block;
+       int err;
+
+       BUG_ON(ac->ac_status != AC_STATUS_FOUND);
+       BUG_ON(ac->ac_b_ex.fe_len <= 0);
+
+       sb = ac->ac_sb;
+       sbi = EXT4_SB(sb);
+       es = sbi->s_es;
+
+       ext4_debug("using block group %lu(%d)\n", ac->ac_b_ex.fe_group,
+                       gdp->bg_free_blocks_count);
+
+       err = -EIO;
+       bitmap_bh = read_block_bitmap(sb, ac->ac_b_ex.fe_group);
+       if (!bitmap_bh)
+               goto out_err;
+
+       err = ext4_journal_get_write_access(handle, bitmap_bh);
+       if (err)
+               goto out_err;
+
+       err = -EIO;
+       gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
+       if (!gdp)
+               goto out_err;
+
+       err = ext4_journal_get_write_access(handle, gdp_bh);
+       if (err)
+               goto out_err;
+
+       block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb)
+               + ac->ac_b_ex.fe_start
+               + le32_to_cpu(es->s_first_data_block);
+
+       if (block == ext4_block_bitmap(sb, gdp) ||
+                       block == ext4_inode_bitmap(sb, gdp) ||
+                       in_range(block, ext4_inode_table(sb, gdp),
+                               EXT4_SB(sb)->s_itb_per_group)) {
+
+               ext4_error(sb, __FUNCTION__,
+                          "Allocating block in system zone - block = %llu",
+                          block);
+       }
+#ifdef AGGRESSIVE_CHECK
+       {
+               int i;
+               for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
+                       BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
+                                               bitmap_bh->b_data));
+               }
+       }
+#endif
+       mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data,
+                               ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
+
+       spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
+       if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
+               gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
+               gdp->bg_free_blocks_count =
+                       cpu_to_le16(ext4_free_blocks_after_init(sb,
+                                               ac->ac_b_ex.fe_group,
+                                               gdp));
+       }
+       gdp->bg_free_blocks_count =
+               cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count)
+                               - ac->ac_b_ex.fe_len);
+       gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
+       spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
+       percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len);
+
+       err = ext4_journal_dirty_metadata(handle, bitmap_bh);
+       if (err)
+               goto out_err;
+       err = ext4_journal_dirty_metadata(handle, gdp_bh);
+
+out_err:
+       sb->s_dirt = 1;
+       put_bh(bitmap_bh);
+       return err;
+}
+
+/*
+ * here we normalize request for locality group
+ * Group request are normalized to s_strip size if we set the same via mount
+ * option. If not we set it to s_mb_group_prealloc which can be configured via
+ * /proc/fs/ext4/<partition>/group_prealloc
+ *
+ * XXX: should we try to preallocate more than the group has now?
+ */
+static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
+{
+       struct super_block *sb = ac->ac_sb;
+       struct ext4_locality_group *lg = ac->ac_lg;
+
+       BUG_ON(lg == NULL);
+       if (EXT4_SB(sb)->s_stripe)
+               ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
+       else
+               ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
+       mb_debug("#%u: goal %lu blocks for locality group\n",
+               current->pid, ac->ac_g_ex.fe_len);
+}
+
+/*
+ * Normalization means making request better in terms of
+ * size and alignment
+ */
+static void ext4_mb_normalize_request(struct ext4_allocation_context *ac,
+                               struct ext4_allocation_request *ar)
+{
+       int bsbits, max;
+       ext4_lblk_t end;
+       struct list_head *cur;
+       loff_t size, orig_size, start_off;
+       ext4_lblk_t start, orig_start;
+       struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
+
+       /* do normalize only data requests, metadata requests
+          do not need preallocation */
+       if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
+               return;
+
+       /* sometime caller may want exact blocks */
+       if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
+               return;
+
+       /* caller may indicate that preallocation isn't
+        * required (it's a tail, for example) */
+       if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
+               return;
+
+       if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
+               ext4_mb_normalize_group_request(ac);
+               return ;
+       }
+
+       bsbits = ac->ac_sb->s_blocksize_bits;
+
+       /* first, let's learn actual file size
+        * given current request is allocated */
+       size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
+       size = size << bsbits;
+       if (size < i_size_read(ac->ac_inode))
+               size = i_size_read(ac->ac_inode);
+
+       /* max available blocks in a free group */
+       max = EXT4_BLOCKS_PER_GROUP(ac->ac_sb) - 1 - 1 -
+                               EXT4_SB(ac->ac_sb)->s_itb_per_group;
+
+#define NRL_CHECK_SIZE(req, size, max,bits)    \
+               (req <= (size) || max <= ((size) >> bits))
+
+       /* first, try to predict filesize */
+       /* XXX: should this table be tunable? */
+       start_off = 0;
+       if (size <= 16 * 1024) {
+               size = 16 * 1024;
+       } else if (size <= 32 * 1024) {
+               size = 32 * 1024;
+       } else if (size <= 64 * 1024) {
+               size = 64 * 1024;
+       } else if (size <= 128 * 1024) {
+               size = 128 * 1024;
+       } else if (size <= 256 * 1024) {
+               size = 256 * 1024;
+       } else if (size <= 512 * 1024) {
+               size = 512 * 1024;
+       } else if (size <= 1024 * 1024) {
+               size = 1024 * 1024;
+       } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, bsbits)) {
+               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
+                                               (20 - bsbits)) << 20;
+               size = 1024 * 1024;
+       } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, bsbits)) {
+               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
+                                                       (22 - bsbits)) << 22;
+               size = 4 * 1024 * 1024;
+       } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
+                                       (8<<20)>>bsbits, max, bsbits)) {
+               start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
+                                                       (23 - bsbits)) << 23;
+               size = 8 * 1024 * 1024;
+       } else {
+               start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
+               size      = ac->ac_o_ex.fe_len << bsbits;
+       }
+       orig_size = size = size >> bsbits;
+       orig_start = start = start_off >> bsbits;
+
+       /* don't cover already allocated blocks in selected range */
+       if (ar->pleft && start <= ar->lleft) {
+               size -= ar->lleft + 1 - start;
+               start = ar->lleft + 1;
+       }
+       if (ar->pright && start + size - 1 >= ar->lright)
+               size -= start + size - ar->lright;
+
+       end = start + size;
+
+       /* check we don't cross already preallocated blocks */
+       rcu_read_lock();
+       list_for_each_rcu(cur, &ei->i_prealloc_list) {
+               struct ext4_prealloc_space *pa;
+               unsigned long pa_end;
+
+               pa = list_entry(cur, struct ext4_prealloc_space, pa_inode_list);
+
+               if (pa->pa_deleted)
+                       continue;
+               spin_lock(&pa->pa_lock);
+               if (pa->pa_deleted) {
+                       spin_unlock(&pa->pa_lock);
+                       continue;
+               }
+
+               pa_end = pa->pa_lstart + pa->pa_len;
+
+               /* PA must not overlap original request */
+               BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
+                       ac->ac_o_ex.fe_logical < pa->pa_lstart));
+
+               /* skip PA normalized request doesn't overlap with */
+               if (pa->pa_lstart >= end) {
+                       spin_unlock(&pa->pa_lock);
+                       continue;
+               }
+               if (pa_end <= start) {
+                       spin_unlock(&pa->pa_lock);
+                       continue;
+               }
+               BUG_ON(pa->pa_lstart <= start && pa_end >= end);
+
+               if (pa_end <= ac->ac_o_ex.fe_logical) {
+                       BUG_ON(pa_end < start);
+                       start = pa_end;
+               }
+
+               if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
+                       BUG_ON(pa->pa_lstart > end);
+                       end = pa->pa_lstart;
+               }
+               spin_unlock(&pa->pa_lock);
+       }
+       rcu_read_unlock();
+       size = end - start;
+
+       /* XXX: extra loop to check we really don't overlap preallocations */
+       rcu_read_lock();
+       list_for_each_rcu(cur, &ei->i_prealloc_list) {
+               struct ext4_prealloc_space *pa;
+               unsigned long pa_end;
+               pa = list_entry(cur, struct ext4_prealloc_space, pa_inode_list);
+               spin_lock(&pa->pa_lock);
+               if (pa->pa_deleted == 0) {
+                       pa_end = pa->pa_lstart + pa->pa_len;
+                       BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
+               }
+               spin_unlock(&pa->pa_lock);
+       }
+       rcu_read_unlock();
+
+       if (start + size <= ac->ac_o_ex.fe_logical &&
+                       start > ac->ac_o_ex.fe_logical) {
+               printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n",
+                       (unsigned long) start, (unsigned long) size,
+                       (unsigned long) ac->ac_o_ex.fe_logical);
+       }
+       BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
+                       start > ac->ac_o_ex.fe_logical);
+       BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
+
+       /* now prepare goal request */
+
+       /* XXX: is it better to align blocks WRT to logical
+        * placement or satisfy big request as is */
+       ac->ac_g_ex.fe_logical = start;
+       ac->ac_g_ex.fe_len = size;
+
+       /* define goal start in order to merge */
+       if (ar->pright && (ar->lright == (start + size))) {
+               /* merge to the right */
+               ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
+                                               &ac->ac_f_ex.fe_group,
+                                               &ac->ac_f_ex.fe_start);
+               ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
+       }
+       if (ar->pleft && (ar->lleft + 1 == start)) {
+               /* merge to the left */
+               ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
+                                               &ac->ac_f_ex.fe_group,
+                                               &ac->ac_f_ex.fe_start);
+               ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
+       }
+
+       mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size,
+               (unsigned) orig_size, (unsigned) start);
+}
+
+static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+
+       if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
+               atomic_inc(&sbi->s_bal_reqs);
+               atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
+               if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len)
+                       atomic_inc(&sbi->s_bal_success);
+               atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
+               if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
+                               ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
+                       atomic_inc(&sbi->s_bal_goals);
+               if (ac->ac_found > sbi->s_mb_max_to_scan)
+                       atomic_inc(&sbi->s_bal_breaks);
+       }
+
+       ext4_mb_store_history(ac);
+}
+
+/*
+ * use blocks preallocated to inode
+ */
+static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
+                               struct ext4_prealloc_space *pa)
+{
+       ext4_fsblk_t start;
+       ext4_fsblk_t end;
+       int len;
+
+       /* found preallocated blocks, use them */
+       start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
+       end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len);
+       len = end - start;
+       ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
+                                       &ac->ac_b_ex.fe_start);
+       ac->ac_b_ex.fe_len = len;
+       ac->ac_status = AC_STATUS_FOUND;
+       ac->ac_pa = pa;
+
+       BUG_ON(start < pa->pa_pstart);
+       BUG_ON(start + len > pa->pa_pstart + pa->pa_len);
+       BUG_ON(pa->pa_free < len);
+       pa->pa_free -= len;
+
+       mb_debug("use %llu/%lu from inode pa %p\n", start, len, pa);
+}
+
+/*
+ * use blocks preallocated to locality group
+ */
+static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
+                               struct ext4_prealloc_space *pa)
+{
+       unsigned len = ac->ac_o_ex.fe_len;
+
+       ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
+                                       &ac->ac_b_ex.fe_group,
+                                       &ac->ac_b_ex.fe_start);
+       ac->ac_b_ex.fe_len = len;
+       ac->ac_status = AC_STATUS_FOUND;
+       ac->ac_pa = pa;
+
+       /* we don't correct pa_pstart or pa_plen here to avoid
+        * possible race when tte group is being loaded concurrently
+        * instead we correct pa later, after blocks are marked
+        * in on-disk bitmap -- see ext4_mb_release_context() */
+       /*
+        * FIXME!! but the other CPUs can look at this particular
+        * pa and think that it have enought free blocks if we
+        * don't update pa_free here right ?
+        */
+       mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
+}
+
+/*
+ * search goal blocks in preallocated space
+ */
+static int ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
+{
+       struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
+       struct ext4_locality_group *lg;
+       struct ext4_prealloc_space *pa;
+       struct list_head *cur;
+
+       /* only data can be preallocated */
+       if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
+               return 0;
+
+       /* first, try per-file preallocation */
+       rcu_read_lock();
+       list_for_each_rcu(cur, &ei->i_prealloc_list) {
+               pa = list_entry(cur, struct ext4_prealloc_space, pa_inode_list);
+
+               /* all fields in this condition don't change,
+                * so we can skip locking for them */
+               if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
+                       ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len)
+                       continue;
+
+               /* found preallocated blocks, use them */
+               spin_lock(&pa->pa_lock);
+               if (pa->pa_deleted == 0 && pa->pa_free) {
+                       atomic_inc(&pa->pa_count);
+                       ext4_mb_use_inode_pa(ac, pa);
+                       spin_unlock(&pa->pa_lock);
+                       ac->ac_criteria = 10;
+                       rcu_read_unlock();
+                       return 1;
+               }
+               spin_unlock(&pa->pa_lock);
+       }
+       rcu_read_unlock();
+
+       /* can we use group allocation? */
+       if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
+               return 0;
+
+       /* inode may have no locality group for some reason */
+       lg = ac->ac_lg;
+       if (lg == NULL)
+               return 0;
+
+       rcu_read_lock();
+       list_for_each_rcu(cur, &lg->lg_prealloc_list) {
+               pa = list_entry(cur, struct ext4_prealloc_space, pa_inode_list);
+               spin_lock(&pa->pa_lock);
+               if (pa->pa_deleted == 0 && pa->pa_free >= ac->ac_o_ex.fe_len) {
+                       atomic_inc(&pa->pa_count);
+                       ext4_mb_use_group_pa(ac, pa);
+                       spin_unlock(&pa->pa_lock);
+                       ac->ac_criteria = 20;
+                       rcu_read_unlock();
+                       return 1;
+               }
+               spin_unlock(&pa->pa_lock);
+       }
+       rcu_read_unlock();
+
+       return 0;
+}
+
+/*
+ * the function goes through all preallocation in this group and marks them
+ * used in in-core bitmap. buddy must be generated from this bitmap
+ * Need to be called with ext4 group lock (ext4_lock_group)
+ */
+static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
+                                       ext4_group_t group)
+{
+       struct ext4_group_info *grp = ext4_get_group_info(sb, group);
+       struct ext4_prealloc_space *pa;
+       struct list_head *cur;
+       ext4_group_t groupnr;
+       ext4_grpblk_t start;
+       int preallocated = 0;
+       int count = 0;
+       int len;
+
+       /* all form of preallocation discards first load group,
+        * so the only competing code is preallocation use.
+        * we don't need any locking here
+        * notice we do NOT ignore preallocations with pa_deleted
+        * otherwise we could leave used blocks available for
+        * allocation in buddy when concurrent ext4_mb_put_pa()
+        * is dropping preallocation
+        */
+       list_for_each(cur, &grp->bb_prealloc_list) {
+               pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
+               spin_lock(&pa->pa_lock);
+               ext4_get_group_no_and_offset(sb, pa->pa_pstart,
+                                            &groupnr, &start);
+               len = pa->pa_len;
+               spin_unlock(&pa->pa_lock);
+               if (unlikely(len == 0))
+                       continue;
+               BUG_ON(groupnr != group);
+               mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
+                                               bitmap, start, len);
+               preallocated += len;
+               count++;
+       }
+       mb_debug("prellocated %u for group %lu\n", preallocated, group);
+}
+
+static void ext4_mb_pa_callback(struct rcu_head *head)
+{
+       struct ext4_prealloc_space *pa;
+       pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
+       kmem_cache_free(ext4_pspace_cachep, pa);
+}
+
+/*
+ * drops a reference to preallocated space descriptor
+ * if this was the last reference and the space is consumed
+ */
+static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
+                       struct super_block *sb, struct ext4_prealloc_space *pa)
+{
+       unsigned long grp;
+
+       if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
+               return;
+
+       /* in this short window concurrent discard can set pa_deleted */
+       spin_lock(&pa->pa_lock);
+       if (pa->pa_deleted == 1) {
+               spin_unlock(&pa->pa_lock);
+               return;
+       }
+
+       pa->pa_deleted = 1;
+       spin_unlock(&pa->pa_lock);
+
+       /* -1 is to protect from crossing allocation group */
+       ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL);
+
+       /*
+        * possible race:
+        *
+        *  P1 (buddy init)                     P2 (regular allocation)
+        *                                      find block B in PA
+        *  copy on-disk bitmap to buddy
+        *                                      mark B in on-disk bitmap
+        *                                      drop PA from group
+        *  mark all PAs in buddy
+        *
+        * thus, P1 initializes buddy with B available. to prevent this
+        * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
+        * against that pair
+        */
+       ext4_lock_group(sb, grp);
+       list_del(&pa->pa_group_list);
+       ext4_unlock_group(sb, grp);
+
+       spin_lock(pa->pa_obj_lock);
+       list_del_rcu(&pa->pa_inode_list);
+       spin_unlock(pa->pa_obj_lock);
+
+       call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
+}
+
+/*
+ * creates new preallocated space for given inode
+ */
+static int ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
+{
+       struct super_block *sb = ac->ac_sb;
+       struct ext4_prealloc_space *pa;
+       struct ext4_group_info *grp;
+       struct ext4_inode_info *ei;
+
+       /* preallocate only when found space is larger then requested */
+       BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
+       BUG_ON(ac->ac_status != AC_STATUS_FOUND);
+       BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
+
+       pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
+       if (pa == NULL)
+               return -ENOMEM;
+
+       if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
+               int winl;
+               int wins;
+               int win;
+               int offs;
+
+               /* we can't allocate as much as normalizer wants.
+                * so, found space must get proper lstart
+                * to cover original request */
+               BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
+               BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
+
+               /* we're limited by original request in that
+                * logical block must be covered any way
+                * winl is window we can move our chunk within */
+               winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
+
+               /* also, we should cover whole original request */
+               wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len;
+
+               /* the smallest one defines real window */
+               win = min(winl, wins);
+
+               offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len;
+               if (offs && offs < win)
+                       win = offs;
+
+               ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win;
+               BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
+               BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
+       }
+
+       /* preallocation can change ac_b_ex, thus we store actually
+        * allocated blocks for history */
+       ac->ac_f_ex = ac->ac_b_ex;
+
+       pa->pa_lstart = ac->ac_b_ex.fe_logical;
+       pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
+       pa->pa_len = ac->ac_b_ex.fe_len;
+       pa->pa_free = pa->pa_len;
+       atomic_set(&pa->pa_count, 1);
+       spin_lock_init(&pa->pa_lock);
+       pa->pa_deleted = 0;
+       pa->pa_linear = 0;
+
+       mb_debug("new inode pa %p: %llu/%u for %u\n", pa,
+                       pa->pa_pstart, pa->pa_len, pa->pa_lstart);
+
+       ext4_mb_use_inode_pa(ac, pa);
+       atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
+
+       ei = EXT4_I(ac->ac_inode);
+       grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
+
+       pa->pa_obj_lock = &ei->i_prealloc_lock;
+       pa->pa_inode = ac->ac_inode;
+
+       ext4_lock_group(sb, ac->ac_b_ex.fe_group);
+       list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
+       ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
+
+       spin_lock(pa->pa_obj_lock);
+       list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
+       spin_unlock(pa->pa_obj_lock);
+
+       return 0;
+}
+
+/*
+ * creates new preallocated space for locality group inodes belongs to
+ */
+static int ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
+{
+       struct super_block *sb = ac->ac_sb;
+       struct ext4_locality_group *lg;
+       struct ext4_prealloc_space *pa;
+       struct ext4_group_info *grp;
+
+       /* preallocate only when found space is larger then requested */
+       BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
+       BUG_ON(ac->ac_status != AC_STATUS_FOUND);
+       BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
+
+       BUG_ON(ext4_pspace_cachep == NULL);
+       pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
+       if (pa == NULL)
+               return -ENOMEM;
+
+       /* preallocation can change ac_b_ex, thus we store actually
+        * allocated blocks for history */
+       ac->ac_f_ex = ac->ac_b_ex;
+
+       pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
+       pa->pa_lstart = pa->pa_pstart;
+       pa->pa_len = ac->ac_b_ex.fe_len;
+       pa->pa_free = pa->pa_len;
+       atomic_set(&pa->pa_count, 1);
+       spin_lock_init(&pa->pa_lock);
+       pa->pa_deleted = 0;
+       pa->pa_linear = 1;
+
+       mb_debug("new group pa %p: %llu/%u for %u\n", pa,
+                       pa->pa_pstart, pa->pa_len, pa->pa_lstart);
+
+       ext4_mb_use_group_pa(ac, pa);
+       atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
+
+       grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
+       lg = ac->ac_lg;
+       BUG_ON(lg == NULL);
+
+       pa->pa_obj_lock = &lg->lg_prealloc_lock;
+       pa->pa_inode = NULL;
+
+       ext4_lock_group(sb, ac->ac_b_ex.fe_group);
+       list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
+       ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
+
+       spin_lock(pa->pa_obj_lock);
+       list_add_tail_rcu(&pa->pa_inode_list, &lg->lg_prealloc_list);
+       spin_unlock(pa->pa_obj_lock);
+
+       return 0;
+}
+
+static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
+{
+       int err;
+
+       if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
+               err = ext4_mb_new_group_pa(ac);
+       else
+               err = ext4_mb_new_inode_pa(ac);
+       return err;
+}
+
+/*
+ * finds all unused blocks in on-disk bitmap, frees them in
+ * in-core bitmap and buddy.
+ * @pa must be unlinked from inode and group lists, so that
+ * nobody else can find/use it.
+ * the caller MUST hold group/inode locks.
+ * TODO: optimize the case when there are no in-core structures yet
+ */
+static int ext4_mb_release_inode_pa(struct ext4_buddy *e4b,
+                               struct buffer_head *bitmap_bh,
+                               struct ext4_prealloc_space *pa)
+{
+       struct ext4_allocation_context ac;
+       struct super_block *sb = e4b->bd_sb;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       unsigned long end;
+       unsigned long next;
+       ext4_group_t group;
+       ext4_grpblk_t bit;
+       sector_t start;
+       int err = 0;
+       int free = 0;
+
+       BUG_ON(pa->pa_deleted == 0);
+       ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
+       BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
+       end = bit + pa->pa_len;
+
+       ac.ac_sb = sb;
+       ac.ac_inode = pa->pa_inode;
+       ac.ac_op = EXT4_MB_HISTORY_DISCARD;
+
+       while (bit < end) {
+               bit = ext4_find_next_zero_bit(bitmap_bh->b_data, end, bit);
+               if (bit >= end)
+                       break;
+               next = ext4_find_next_bit(bitmap_bh->b_data, end, bit);
+               if (next > end)
+                       next = end;
+               start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit +
+                               le32_to_cpu(sbi->s_es->s_first_data_block);
+               mb_debug("    free preallocated %u/%u in group %u\n",
+                               (unsigned) start, (unsigned) next - bit,
+                               (unsigned) group);
+               free += next - bit;
+
+               ac.ac_b_ex.fe_group = group;
+               ac.ac_b_ex.fe_start = bit;
+               ac.ac_b_ex.fe_len = next - bit;
+               ac.ac_b_ex.fe_logical = 0;
+               ext4_mb_store_history(&ac);
+
+               mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
+               bit = next + 1;
+       }
+       if (free != pa->pa_free) {
+               printk(KERN_ERR "pa %p: logic %lu, phys. %lu, len %lu\n",
+                       pa, (unsigned long) pa->pa_lstart,
+                       (unsigned long) pa->pa_pstart,
+                       (unsigned long) pa->pa_len);
+               printk(KERN_ERR "free %u, pa_free %u\n", free, pa->pa_free);
+       }
+       BUG_ON(free != pa->pa_free);
+       atomic_add(free, &sbi->s_mb_discarded);
+
+       return err;
+}
+
+static int ext4_mb_release_group_pa(struct ext4_buddy *e4b,
+                               struct ext4_prealloc_space *pa)
+{
+       struct ext4_allocation_context ac;
+       struct super_block *sb = e4b->bd_sb;
+       ext4_group_t group;
+       ext4_grpblk_t bit;
+
+       ac.ac_op = EXT4_MB_HISTORY_DISCARD;
+
+       BUG_ON(pa->pa_deleted == 0);
+       ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
+       BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
+       mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
+       atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
+
+       ac.ac_sb = sb;
+       ac.ac_inode = NULL;
+       ac.ac_b_ex.fe_group = group;
+       ac.ac_b_ex.fe_start = bit;
+       ac.ac_b_ex.fe_len = pa->pa_len;
+       ac.ac_b_ex.fe_logical = 0;
+       ext4_mb_store_history(&ac);
+
+       return 0;
+}
+
+/*
+ * releases all preallocations in given group
+ *
+ * first, we need to decide discard policy:
+ * - when do we discard
+ *   1) ENOSPC
+ * - how many do we discard
+ *   1) how many requested
+ */
+static int ext4_mb_discard_group_preallocations(struct super_block *sb,
+                                       ext4_group_t group, int needed)
+{
+       struct ext4_group_info *grp = ext4_get_group_info(sb, group);
+       struct buffer_head *bitmap_bh = NULL;
+       struct ext4_prealloc_space *pa, *tmp;
+       struct list_head list;
+       struct ext4_buddy e4b;
+       int err;
+       int busy = 0;
+       int free = 0;
+
+       mb_debug("discard preallocation for group %lu\n", group);
+
+       if (list_empty(&grp->bb_prealloc_list))
+               return 0;
+
+       bitmap_bh = read_block_bitmap(sb, group);
+       if (bitmap_bh == NULL) {
+               /* error handling here */
+               ext4_mb_release_desc(&e4b);
+               BUG_ON(bitmap_bh == NULL);
+       }
+
+       err = ext4_mb_load_buddy(sb, group, &e4b);
+       BUG_ON(err != 0); /* error handling here */
+
+       if (needed == 0)
+               needed = EXT4_BLOCKS_PER_GROUP(sb) + 1;
+
+       grp = ext4_get_group_info(sb, group);
+       INIT_LIST_HEAD(&list);
+
+repeat:
+       ext4_lock_group(sb, group);
+       list_for_each_entry_safe(pa, tmp,
+                               &grp->bb_prealloc_list, pa_group_list) {
+               spin_lock(&pa->pa_lock);
+               if (atomic_read(&pa->pa_count)) {
+                       spin_unlock(&pa->pa_lock);
+                       busy = 1;
+                       continue;
+               }
+               if (pa->pa_deleted) {
+                       spin_unlock(&pa->pa_lock);
+                       continue;
+               }
+
+               /* seems this one can be freed ... */
+               pa->pa_deleted = 1;
+
+               /* we can trust pa_free ... */
+               free += pa->pa_free;
+
+               spin_unlock(&pa->pa_lock);
+
+               list_del(&pa->pa_group_list);
+               list_add(&pa->u.pa_tmp_list, &list);
+       }
+
+       /* if we still need more blocks and some PAs were used, try again */
+       if (free < needed && busy) {
+               busy = 0;
+               ext4_unlock_group(sb, group);
+               /*
+                * Yield the CPU here so that we don't get soft lockup
+                * in non preempt case.
+                */
+               yield();
+               goto repeat;
+       }
+
+       /* found anything to free? */
+       if (list_empty(&list)) {
+               BUG_ON(free != 0);
+               goto out;
+       }
+
+       /* now free all selected PAs */
+       list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
+
+               /* remove from object (inode or locality group) */
+               spin_lock(pa->pa_obj_lock);
+               list_del_rcu(&pa->pa_inode_list);
+               spin_unlock(pa->pa_obj_lock);
+
+               if (pa->pa_linear)
+                       ext4_mb_release_group_pa(&e4b, pa);
+               else
+                       ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
+
+               list_del(&pa->u.pa_tmp_list);
+               call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
+       }
+
+out:
+       ext4_unlock_group(sb, group);
+       ext4_mb_release_desc(&e4b);
+       put_bh(bitmap_bh);
+       return free;
+}
+
+/*
+ * releases all non-used preallocated blocks for given inode
+ *
+ * It's important to discard preallocations under i_data_sem
+ * We don't want another block to be served from the prealloc
+ * space when we are discarding the inode prealloc space.
+ *
+ * FIXME!! Make sure it is valid at all the call sites
+ */
+void ext4_mb_discard_inode_preallocations(struct inode *inode)
+{
+       struct ext4_inode_info *ei = EXT4_I(inode);
+       struct super_block *sb = inode->i_sb;
+       struct buffer_head *bitmap_bh = NULL;
+       struct ext4_prealloc_space *pa, *tmp;
+       ext4_group_t group = 0;
+       struct list_head list;
+       struct ext4_buddy e4b;
+       int err;
+
+       if (!test_opt(sb, MBALLOC) || !S_ISREG(inode->i_mode)) {
+               /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
+               return;
+       }
+
+       mb_debug("discard preallocation for inode %lu\n", inode->i_ino);
+
+       INIT_LIST_HEAD(&list);
+
+repeat:
+       /* first, collect all pa's in the inode */
+       spin_lock(&ei->i_prealloc_lock);
+       while (!list_empty(&ei->i_prealloc_list)) {
+               pa = list_entry(ei->i_prealloc_list.next,
+                               struct ext4_prealloc_space, pa_inode_list);
+               BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
+               spin_lock(&pa->pa_lock);
+               if (atomic_read(&pa->pa_count)) {
+                       /* this shouldn't happen often - nobody should
+                        * use preallocation while we're discarding it */
+                       spin_unlock(&pa->pa_lock);
+                       spin_unlock(&ei->i_prealloc_lock);
+                       printk(KERN_ERR "uh-oh! used pa while discarding\n");
+                       WARN_ON(1);
+                       schedule_timeout_uninterruptible(HZ);
+                       goto repeat;
+
+               }
+               if (pa->pa_deleted == 0) {
+                       pa->pa_deleted = 1;
+                       spin_unlock(&pa->pa_lock);
+                       list_del_rcu(&pa->pa_inode_list);
+                       list_add(&pa->u.pa_tmp_list, &list);
+                       continue;
+               }
+
+               /* someone is deleting pa right now */
+               spin_unlock(&pa->pa_lock);
+               spin_unlock(&ei->i_prealloc_lock);
+
+               /* we have to wait here because pa_deleted
+                * doesn't mean pa is already unlinked from
+                * the list. as we might be called from
+                * ->clear_inode() the inode will get freed
+                * and concurrent thread which is unlinking
+                * pa from inode's list may access already
+                * freed memory, bad-bad-bad */
+
+               /* XXX: if this happens too often, we can
+                * add a flag to force wait only in case
+                * of ->clear_inode(), but not in case of
+                * regular truncate */
+               schedule_timeout_uninterruptible(HZ);
+               goto repeat;
+       }
+       spin_unlock(&ei->i_prealloc_lock);
+
+       list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
+               BUG_ON(pa->pa_linear != 0);
+               ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
+
+               err = ext4_mb_load_buddy(sb, group, &e4b);
+               BUG_ON(err != 0); /* error handling here */
+
+               bitmap_bh = read_block_bitmap(sb, group);
+               if (bitmap_bh == NULL) {
+                       /* error handling here */
+                       ext4_mb_release_desc(&e4b);
+                       BUG_ON(bitmap_bh == NULL);
+               }
+
+               ext4_lock_group(sb, group);
+               list_del(&pa->pa_group_list);
+               ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
+               ext4_unlock_group(sb, group);
+
+               ext4_mb_release_desc(&e4b);
+               put_bh(bitmap_bh);
+
+               list_del(&pa->u.pa_tmp_list);
+               call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
+       }
+}
+
+/*
+ * finds all preallocated spaces and return blocks being freed to them
+ * if preallocated space becomes full (no block is used from the space)
+ * then the function frees space in buddy
+ * XXX: at the moment, truncate (which is the only way to free blocks)
+ * discards all preallocations
+ */
+static void ext4_mb_return_to_preallocation(struct inode *inode,
+                                       struct ext4_buddy *e4b,
+                                       sector_t block, int count)
+{
+       BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list));
+}
+#ifdef MB_DEBUG
+static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
+{
+       struct super_block *sb = ac->ac_sb;
+       ext4_group_t i;
+
+       printk(KERN_ERR "EXT4-fs: Can't allocate:"
+                       " Allocation context details:\n");
+       printk(KERN_ERR "EXT4-fs: status %d flags %d\n",
+                       ac->ac_status, ac->ac_flags);
+       printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, "
+                       "best %lu/%lu/%lu@%lu cr %d\n",
+                       (unsigned long)ac->ac_o_ex.fe_group,
+                       (unsigned long)ac->ac_o_ex.fe_start,
+                       (unsigned long)ac->ac_o_ex.fe_len,
+                       (unsigned long)ac->ac_o_ex.fe_logical,
+                       (unsigned long)ac->ac_g_ex.fe_group,
+                       (unsigned long)ac->ac_g_ex.fe_start,
+                       (unsigned long)ac->ac_g_ex.fe_len,
+                       (unsigned long)ac->ac_g_ex.fe_logical,
+                       (unsigned long)ac->ac_b_ex.fe_group,
+                       (unsigned long)ac->ac_b_ex.fe_start,
+                       (unsigned long)ac->ac_b_ex.fe_len,
+                       (unsigned long)ac->ac_b_ex.fe_logical,
+                       (int)ac->ac_criteria);
+       printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned,
+               ac->ac_found);
+       printk(KERN_ERR "EXT4-fs: groups: \n");
+       for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
+               struct ext4_group_info *grp = ext4_get_group_info(sb, i);
+               struct ext4_prealloc_space *pa;
+               ext4_grpblk_t start;
+               struct list_head *cur;
+               ext4_lock_group(sb, i);
+               list_for_each(cur, &grp->bb_prealloc_list) {
+                       pa = list_entry(cur, struct ext4_prealloc_space,
+                                       pa_group_list);
+                       spin_lock(&pa->pa_lock);
+                       ext4_get_group_no_and_offset(sb, pa->pa_pstart,
+                                                    NULL, &start);
+                       spin_unlock(&pa->pa_lock);
+                       printk(KERN_ERR "PA:%lu:%d:%u \n", i,
+                                                       start, pa->pa_len);
+               }
+               ext4_lock_group(sb, i);
+
+               if (grp->bb_free == 0)
+                       continue;
+               printk(KERN_ERR "%lu: %d/%d \n",
+                      i, grp->bb_free, grp->bb_fragments);
+       }
+       printk(KERN_ERR "\n");
+}
+#else
+static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
+{
+       return;
+}
+#endif
+
+/*
+ * We use locality group preallocation for small size file. The size of the
+ * file is determined by the current size or the resulting size after
+ * allocation which ever is larger
+ *
+ * One can tune this size via /proc/fs/ext4/<partition>/stream_req
+ */
+static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+       int bsbits = ac->ac_sb->s_blocksize_bits;
+       loff_t size, isize;
+
+       if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
+               return;
+
+       size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len;
+       isize = i_size_read(ac->ac_inode) >> bsbits;
+       size = max(size, isize);
+
+       /* don't use group allocation for large files */
+       if (size >= sbi->s_mb_stream_request)
+               return;
+
+       if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
+               return;
+
+       BUG_ON(ac->ac_lg != NULL);
+       /*
+        * locality group prealloc space are per cpu. The reason for having
+        * per cpu locality group is to reduce the contention between block
+        * request from multiple CPUs.
+        */
+       ac->ac_lg = &sbi->s_locality_groups[get_cpu()];
+       put_cpu();
+
+       /* we're going to use group allocation */
+       ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
+
+       /* serialize all allocations in the group */
+       mutex_lock(&ac->ac_lg->lg_mutex);
+}
+
+static int ext4_mb_initialize_context(struct ext4_allocation_context *ac,
+                               struct ext4_allocation_request *ar)
+{
+       struct super_block *sb = ar->inode->i_sb;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       struct ext4_super_block *es = sbi->s_es;
+       ext4_group_t group;
+       unsigned long len;
+       unsigned long goal;
+       ext4_grpblk_t block;
+
+       /* we can't allocate > group size */
+       len = ar->len;
+
+       /* just a dirty hack to filter too big requests  */
+       if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10)
+               len = EXT4_BLOCKS_PER_GROUP(sb) - 10;
+
+       /* start searching from the goal */
+       goal = ar->goal;
+       if (goal < le32_to_cpu(es->s_first_data_block) ||
+                       goal >= ext4_blocks_count(es))
+               goal = le32_to_cpu(es->s_first_data_block);
+       ext4_get_group_no_and_offset(sb, goal, &group, &block);
+
+       /* set up allocation goals */
+       ac->ac_b_ex.fe_logical = ar->logical;
+       ac->ac_b_ex.fe_group = 0;
+       ac->ac_b_ex.fe_start = 0;
+       ac->ac_b_ex.fe_len = 0;
+       ac->ac_status = AC_STATUS_CONTINUE;
+       ac->ac_groups_scanned = 0;
+       ac->ac_ex_scanned = 0;
+       ac->ac_found = 0;
+       ac->ac_sb = sb;
+       ac->ac_inode = ar->inode;
+       ac->ac_o_ex.fe_logical = ar->logical;
+       ac->ac_o_ex.fe_group = group;
+       ac->ac_o_ex.fe_start = block;
+       ac->ac_o_ex.fe_len = len;
+       ac->ac_g_ex.fe_logical = ar->logical;
+       ac->ac_g_ex.fe_group = group;
+       ac->ac_g_ex.fe_start = block;
+       ac->ac_g_ex.fe_len = len;
+       ac->ac_f_ex.fe_len = 0;
+       ac->ac_flags = ar->flags;
+       ac->ac_2order = 0;
+       ac->ac_criteria = 0;
+       ac->ac_pa = NULL;
+       ac->ac_bitmap_page = NULL;
+       ac->ac_buddy_page = NULL;
+       ac->ac_lg = NULL;
+
+       /* we have to define context: we'll we work with a file or
+        * locality group. this is a policy, actually */
+       ext4_mb_group_or_file(ac);
+
+       mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
+                       "left: %u/%u, right %u/%u to %swritable\n",
+                       (unsigned) ar->len, (unsigned) ar->logical,
+                       (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
+                       (unsigned) ar->lleft, (unsigned) ar->pleft,
+                       (unsigned) ar->lright, (unsigned) ar->pright,
+                       atomic_read(&ar->inode->i_writecount) ? "" : "non-");
+       return 0;
+
+}
+
+/*
+ * release all resource we used in allocation
+ */
+static int ext4_mb_release_context(struct ext4_allocation_context *ac)
+{
+       if (ac->ac_pa) {
+               if (ac->ac_pa->pa_linear) {
+                       /* see comment in ext4_mb_use_group_pa() */
+                       spin_lock(&ac->ac_pa->pa_lock);
+                       ac->ac_pa->pa_pstart += ac->ac_b_ex.fe_len;
+                       ac->ac_pa->pa_lstart += ac->ac_b_ex.fe_len;
+                       ac->ac_pa->pa_free -= ac->ac_b_ex.fe_len;
+                       ac->ac_pa->pa_len -= ac->ac_b_ex.fe_len;
+                       spin_unlock(&ac->ac_pa->pa_lock);
+               }
+               ext4_mb_put_pa(ac, ac->ac_sb, ac->ac_pa);
+       }
+       if (ac->ac_bitmap_page)
+               page_cache_release(ac->ac_bitmap_page);
+       if (ac->ac_buddy_page)
+               page_cache_release(ac->ac_buddy_page);
+       if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
+               mutex_unlock(&ac->ac_lg->lg_mutex);
+       ext4_mb_collect_stats(ac);
+       return 0;
+}
+
+static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
+{
+       ext4_group_t i;
+       int ret;
+       int freed = 0;
+
+       for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) {
+               ret = ext4_mb_discard_group_preallocations(sb, i, needed);
+               freed += ret;
+               needed -= ret;
+       }
+
+       return freed;
+}
+
+/*
+ * Main entry point into mballoc to allocate blocks
+ * it tries to use preallocation first, then falls back
+ * to usual allocation
+ */
+ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
+                                struct ext4_allocation_request *ar, int *errp)
+{
+       struct ext4_allocation_context ac;
+       struct ext4_sb_info *sbi;
+       struct super_block *sb;
+       ext4_fsblk_t block = 0;
+       int freed;
+       int inquota;
+
+       sb = ar->inode->i_sb;
+       sbi = EXT4_SB(sb);
+
+       if (!test_opt(sb, MBALLOC)) {
+               block = ext4_new_blocks_old(handle, ar->inode, ar->goal,
+                                           &(ar->len), errp);
+               return block;
+       }
+
+       while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) {
+               ar->flags |= EXT4_MB_HINT_NOPREALLOC;
+               ar->len--;
+       }
+       if (ar->len == 0) {
+               *errp = -EDQUOT;
+               return 0;
+       }
+       inquota = ar->len;
+
+       ext4_mb_poll_new_transaction(sb, handle);
+
+       *errp = ext4_mb_initialize_context(&ac, ar);
+       if (*errp) {
+               ar->len = 0;
+               goto out;
+       }
+
+       ac.ac_op = EXT4_MB_HISTORY_PREALLOC;
+       if (!ext4_mb_use_preallocated(&ac)) {
+
+               ac.ac_op = EXT4_MB_HISTORY_ALLOC;
+               ext4_mb_normalize_request(&ac, ar);
+
+repeat:
+               /* allocate space in core */
+               ext4_mb_regular_allocator(&ac);
+
+               /* as we've just preallocated more space than
+                * user requested orinally, we store allocated
+                * space in a special descriptor */
+               if (ac.ac_status == AC_STATUS_FOUND &&
+                               ac.ac_o_ex.fe_len < ac.ac_b_ex.fe_len)
+                       ext4_mb_new_preallocation(&ac);
+       }
+
+       if (likely(ac.ac_status == AC_STATUS_FOUND)) {
+               ext4_mb_mark_diskspace_used(&ac, handle);
+               *errp = 0;
+               block = ext4_grp_offs_to_block(sb, &ac.ac_b_ex);
+               ar->len = ac.ac_b_ex.fe_len;
+       } else {
+               freed  = ext4_mb_discard_preallocations(sb, ac.ac_o_ex.fe_len);
+               if (freed)
+                       goto repeat;
+               *errp = -ENOSPC;
+               ac.ac_b_ex.fe_len = 0;
+               ar->len = 0;
+               ext4_mb_show_ac(&ac);
+       }
+
+       ext4_mb_release_context(&ac);
+
+out:
+       if (ar->len < inquota)
+               DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len);
+
+       return block;
+}
+static void ext4_mb_poll_new_transaction(struct super_block *sb,
+                                               handle_t *handle)
+{
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+
+       if (sbi->s_last_transaction == handle->h_transaction->t_tid)
+               return;
+
+       /* new transaction! time to close last one and free blocks for
+        * committed transaction. we know that only transaction can be
+        * active, so previos transaction can be being logged and we
+        * know that transaction before previous is known to be already
+        * logged. this means that now we may free blocks freed in all
+        * transactions before previous one. hope I'm clear enough ... */
+
+       spin_lock(&sbi->s_md_lock);
+       if (sbi->s_last_transaction != handle->h_transaction->t_tid) {
+               mb_debug("new transaction %lu, old %lu\n",
+                               (unsigned long) handle->h_transaction->t_tid,
+                               (unsigned long) sbi->s_last_transaction);
+               list_splice_init(&sbi->s_closed_transaction,
+                               &sbi->s_committed_transaction);
+               list_splice_init(&sbi->s_active_transaction,
+                               &sbi->s_closed_transaction);
+               sbi->s_last_transaction = handle->h_transaction->t_tid;
+       }
+       spin_unlock(&sbi->s_md_lock);
+
+       ext4_mb_free_committed_blocks(sb);
+}
+
+static int ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
+                         ext4_group_t group, ext4_grpblk_t block, int count)
+{
+       struct ext4_group_info *db = e4b->bd_info;
+       struct super_block *sb = e4b->bd_sb;
+       struct ext4_sb_info *sbi = EXT4_SB(sb);
+       struct ext4_free_metadata *md;
+       int i;
+
+       BUG_ON(e4b->bd_bitmap_page == NULL);
+       BUG_ON(e4b->bd_buddy_page == NULL);
+
+       ext4_lock_group(sb, group);
+       for (i = 0; i < count; i++) {
+               md = db->bb_md_cur;
+               if (md && db->bb_tid != handle->h_transaction->t_tid) {
+                       db->bb_md_cur = NULL;
+                       md = NULL;
+               }
+
+               if (md == NULL) {
+                       ext4_unlock_group(sb, group);
+                       md = kmalloc(sizeof(*md), GFP_NOFS);
+                       if (md == NULL)
+                               return -ENOMEM;
+                       md->num = 0;
+                       md->group = group;
+
+                       ext4_lock_group(sb, group);
+                       if (db->bb_md_cur == NULL) {
+                               spin_lock(&sbi->s_md_lock);
+                               list_add(&md->list, &sbi->s_active_transaction);
+                               spin_unlock(&sbi->s_md_lock);
+                               /* protect buddy cache from being freed,
+                                * otherwise we'll refresh it from
+                                * on-disk bitmap and lose not-yet-available
+                                * blocks */
+                               page_cache_get(e4b->bd_buddy_page);
+                               page_cache_get(e4b->bd_bitmap_page);
+                               db->bb_md_cur = md;
+                               db->bb_tid = handle->h_transaction->t_tid;
+                               mb_debug("new md 0x%p for group %lu\n",
+                                               md, md->group);
+                       } else {
+                               kfree(md);
+                               md = db->bb_md_cur;
+                       }
+               }
+
+               BUG_ON(md->num >= EXT4_BB_MAX_BLOCKS);
+               md->blocks[md->num] = block + i;
+               md->num++;
+               if (md->num == EXT4_BB_MAX_BLOCKS) {
+                       /* no more space, put full container on a sb's list */
+                       db->bb_md_cur = NULL;
+               }
+       }
+       ext4_unlock_group(sb, group);
+       return 0;
+}
+
+/*
+ * Main entry point into mballoc to free blocks
+ */
+void ext4_mb_free_blocks(handle_t *handle, struct inode *inode,
+                       unsigned long block, unsigned long count,
+                       int metadata, unsigned long *freed)
+{
+       struct buffer_head *bitmap_bh = 0;
+       struct super_block *sb = inode->i_sb;
+       struct ext4_allocation_context ac;
+       struct ext4_group_desc *gdp;
+       struct ext4_super_block *es;
+       unsigned long overflow;
+       ext4_grpblk_t bit;
+       struct buffer_head *gd_bh;
+       ext4_group_t block_group;
+       struct ext4_sb_info *sbi;
+       struct ext4_buddy e4b;
+       int err = 0;
+       int ret;
+
+       *freed = 0;
+
+       ext4_mb_poll_new_transaction(sb, handle);
+
+       sbi = EXT4_SB(sb);
+       es = EXT4_SB(sb)->s_es;
+       if (block < le32_to_cpu(es->s_first_data_block) ||
+           block + count < block ||
+           block + count > ext4_blocks_count(es)) {
+               ext4_error(sb, __FUNCTION__,
+                           "Freeing blocks not in datazone - "
+                           "block = %lu, count = %lu", block, count);
+               goto error_return;
+       }
+
+       ext4_debug("freeing block %lu\n", block);
+
+       ac.ac_op = EXT4_MB_HISTORY_FREE;
+       ac.ac_inode = inode;
+       ac.ac_sb = sb;
+
+do_more:
+       overflow = 0;
+       ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
+
+       /*
+        * Check to see if we are freeing blocks across a group
+        * boundary.
+        */
+       if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
+               overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb);
+               count -= overflow;
+       }
+       bitmap_bh = read_block_bitmap(sb, block_group);
+       if (!bitmap_bh)
+               goto error_return;
+       gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
+       if (!gdp)
+               goto error_return;
+
+       if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
+           in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
+           in_range(block, ext4_inode_table(sb, gdp),
+                     EXT4_SB(sb)->s_itb_per_group) ||
+           in_range(block + count - 1, ext4_inode_table(sb, gdp),
+                     EXT4_SB(sb)->s_itb_per_group)) {
+
+               ext4_error(sb, __FUNCTION__,
+                          "Freeing blocks in system zone - "
+                          "Block = %lu, count = %lu", block, count);
+       }
+
+       BUFFER_TRACE(bitmap_bh, "getting write access");
+       err = ext4_journal_get_write_access(handle, bitmap_bh);
+       if (err)
+               goto error_return;
+
+       /*
+        * We are about to modify some metadata.  Call the journal APIs
+        * to unshare ->b_data if a currently-committing transaction is
+        * using it
+        */
+       BUFFER_TRACE(gd_bh, "get_write_access");
+       err = ext4_journal_get_write_access(handle, gd_bh);
+       if (err)
+               goto error_return;
+
+       err = ext4_mb_load_buddy(sb, block_group, &e4b);
+       if (err)
+               goto error_return;
+
+#ifdef AGGRESSIVE_CHECK
+       {
+               int i;
+               for (i = 0; i < count; i++)
+                       BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
+       }
+#endif
+       mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data,
+                       bit, count);
+
+       /* We dirtied the bitmap block */
+       BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
+       err = ext4_journal_dirty_metadata(handle, bitmap_bh);
+
+       ac.ac_b_ex.fe_group = block_group;
+       ac.ac_b_ex.fe_start = bit;
+       ac.ac_b_ex.fe_len = count;
+       ext4_mb_store_history(&ac);
+
+       if (metadata) {
+               /* blocks being freed are metadata. these blocks shouldn't
+                * be used until this transaction is committed */
+               ext4_mb_free_metadata(handle, &e4b, block_group, bit, count);
+       } else {
+               ext4_lock_group(sb, block_group);
+               err = mb_free_blocks(inode, &e4b, bit, count);
+               ext4_mb_return_to_preallocation(inode, &e4b, block, count);
+               ext4_unlock_group(sb, block_group);
+               BUG_ON(err != 0);
+       }
+
+       spin_lock(sb_bgl_lock(sbi, block_group));
+       gdp->bg_free_blocks_count =
+               cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) + count);
+       gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp);
+       spin_unlock(sb_bgl_lock(sbi, block_group));
+       percpu_counter_add(&sbi->s_freeblocks_counter, count);
+
+       ext4_mb_release_desc(&e4b);
+
+       *freed += count;
+
+       /* And the group descriptor block */
+       BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
+       ret = ext4_journal_dirty_metadata(handle, gd_bh);
+       if (!err)
+               err = ret;
+
+       if (overflow && !err) {
+               block += count;
+               count = overflow;
+               put_bh(bitmap_bh);
+               goto do_more;
+       }
+       sb->s_dirt = 1;
+error_return:
+       brelse(bitmap_bh);
+       ext4_std_error(sb, err);
+       return;
+}
index ec7cb56..3ebc233 100644 (file)
@@ -236,10 +236,10 @@ static int free_dind_blocks(handle_t *handle,
        for (i = 0; i < max_entries; i++) {
                if (tmp_idata[i])
                        ext4_free_blocks(handle, inode,
-                                       le32_to_cpu(tmp_idata[i]), 1);
+                                       le32_to_cpu(tmp_idata[i]), 1, 1);
        }
        put_bh(bh);
-       ext4_free_blocks(handle, inode, le32_to_cpu(i_data), 1);
+       ext4_free_blocks(handle, inode, le32_to_cpu(i_data), 1, 1);
        return 0;
 }
 
@@ -267,7 +267,7 @@ static int free_tind_blocks(handle_t *handle,
                }
        }
        put_bh(bh);
-       ext4_free_blocks(handle, inode, le32_to_cpu(i_data), 1);
+       ext4_free_blocks(handle, inode, le32_to_cpu(i_data), 1, 1);
        return 0;
 }
 
@@ -278,7 +278,7 @@ static int free_ind_block(handle_t *handle, struct inode *inode)
 
        if (ei->i_data[EXT4_IND_BLOCK])
                ext4_free_blocks(handle, inode,
-                               le32_to_cpu(ei->i_data[EXT4_IND_BLOCK]), 1);
+                               le32_to_cpu(ei->i_data[EXT4_IND_BLOCK]), 1, 1);
 
        if (ei->i_data[EXT4_DIND_BLOCK]) {
                retval = free_dind_blocks(handle, inode,
@@ -365,7 +365,7 @@ static int free_ext_idx(handle_t *handle, struct inode *inode,
                }
        }
        put_bh(bh);
-       ext4_free_blocks(handle, inode, block, 1);
+       ext4_free_blocks(handle, inode, block, 1, 1);
        return retval;
 }
 
index 64fc7f1..3a51ffc 100644 (file)
@@ -503,6 +503,7 @@ static void ext4_put_super (struct super_block * sb)
        struct ext4_super_block *es = sbi->s_es;
        int i;
 
+       ext4_mb_release(sb);
        ext4_ext_release(sb);
        ext4_xattr_put_super(sb);
        jbd2_journal_destroy(sbi->s_journal);
@@ -569,6 +570,8 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
        ei->i_block_alloc_info = NULL;
        ei->vfs_inode.i_version = 1;
        memset(&ei->i_cached_extent, 0, sizeof(struct ext4_ext_cache));
+       INIT_LIST_HEAD(&ei->i_prealloc_list);
+       spin_lock_init(&ei->i_prealloc_lock);
        return &ei->vfs_inode;
 }
 
@@ -881,6 +884,7 @@ enum {
        Opt_jqfmt_vfsold, Opt_jqfmt_vfsv0, Opt_quota, Opt_noquota,
        Opt_ignore, Opt_barrier, Opt_err, Opt_resize, Opt_usrquota,
        Opt_grpquota, Opt_extents, Opt_noextents, Opt_i_version,
+       Opt_mballoc, Opt_nomballoc, Opt_stripe,
 };
 
 static match_table_t tokens = {
@@ -935,6 +939,9 @@ static match_table_t tokens = {
        {Opt_extents, "extents"},
        {Opt_noextents, "noextents"},
        {Opt_i_version, "i_version"},
+       {Opt_mballoc, "mballoc"},
+       {Opt_nomballoc, "nomballoc"},
+       {Opt_stripe, "stripe=%u"},
        {Opt_err, NULL},
        {Opt_resize, "resize"},
 };
@@ -1284,6 +1291,19 @@ clear_qf_name:
                        set_opt(sbi->s_mount_opt, I_VERSION);
                        sb->s_flags |= MS_I_VERSION;
                        break;
+               case Opt_mballoc:
+                       set_opt(sbi->s_mount_opt, MBALLOC);
+                       break;
+               case Opt_nomballoc:
+                       clear_opt(sbi->s_mount_opt, MBALLOC);
+                       break;
+               case Opt_stripe:
+                       if (match_int(&args[0], &option))
+                               return 0;
+                       if (option < 0)
+                               return 0;
+                       sbi->s_stripe = option;
+                       break;
                default:
                        printk (KERN_ERR
                                "EXT4-fs: Unrecognized mount option \"%s\" "
@@ -1742,6 +1762,34 @@ static ext4_fsblk_t descriptor_loc(struct super_block *sb,
        return (has_super + ext4_group_first_block_no(sb, bg));
 }
 
+/**
+ * ext4_get_stripe_size: Get the stripe size.
+ * @sbi: In memory super block info
+ *
+ * If we have specified it via mount option, then
+ * use the mount option value. If the value specified at mount time is
+ * greater than the blocks per group use the super block value.
+ * If the super block value is greater than blocks per group return 0.
+ * Allocator needs it be less than blocks per group.
+ *
+ */
+static unsigned long ext4_get_stripe_size(struct ext4_sb_info *sbi)
+{
+       unsigned long stride = le16_to_cpu(sbi->s_es->s_raid_stride);
+       unsigned long stripe_width =
+                       le32_to_cpu(sbi->s_es->s_raid_stripe_width);
+
+       if (sbi->s_stripe && sbi->s_stripe <= sbi->s_blocks_per_group)
+               return sbi->s_stripe;
+
+       if (stripe_width <= sbi->s_blocks_per_group)
+               return stripe_width;
+
+       if (stride <= sbi->s_blocks_per_group)
+               return stride;
+
+       return 0;
+}
 
 static int ext4_fill_super (struct super_block *sb, void *data, int silent)
                                __releases(kernel_sem)
@@ -2091,6 +2139,8 @@ static int ext4_fill_super (struct super_block *sb, void *data, int silent)
        sbi->s_rsv_window_head.rsv_goal_size = 0;
        ext4_rsv_window_add(sb, &sbi->s_rsv_window_head);
 
+       sbi->s_stripe = ext4_get_stripe_size(sbi);
+
        /*
         * set up enough so that it can read an inode
         */
@@ -2250,6 +2300,7 @@ static int ext4_fill_super (struct super_block *sb, void *data, int silent)
                "writeback");
 
        ext4_ext_init(sb);
+       ext4_mb_init(sb, needs_recovery);
 
        lock_kernel();
        return 0;
@@ -3232,9 +3283,15 @@ static struct file_system_type ext4dev_fs_type = {
 
 static int __init init_ext4_fs(void)
 {
-       int err = init_ext4_xattr();
+       int err;
+
+       err = init_ext4_mballoc();
        if (err)
                return err;
+
+       err = init_ext4_xattr();
+       if (err)
+               goto out2;
        err = init_inodecache();
        if (err)
                goto out1;
@@ -3246,6 +3303,8 @@ out:
        destroy_inodecache();
 out1:
        exit_ext4_xattr();
+out2:
+       exit_ext4_mballoc();
        return err;
 }
 
@@ -3254,6 +3313,7 @@ static void __exit exit_ext4_fs(void)
        unregister_filesystem(&ext4dev_fs_type);
        destroy_inodecache();
        exit_ext4_xattr();
+       exit_ext4_mballoc();
 }
 
 MODULE_AUTHOR("Remy Card, Stephen Tweedie, Andrew Morton, Andreas Dilger, Theodore Ts'o and others");
index 8638730..d796213 100644 (file)
@@ -480,7 +480,7 @@ ext4_xattr_release_block(handle_t *handle, struct inode *inode,
                ea_bdebug(bh, "refcount now=0; freeing");
                if (ce)
                        mb_cache_entry_free(ce);
-               ext4_free_blocks(handle, inode, bh->b_blocknr, 1);
+               ext4_free_blocks(handle, inode, bh->b_blocknr, 1, 1);
                get_bh(bh);
                ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
        } else {
@@ -821,7 +821,7 @@ inserted:
                        new_bh = sb_getblk(sb, block);
                        if (!new_bh) {
 getblk_failed:
-                               ext4_free_blocks(handle, inode, block, 1);
+                               ext4_free_blocks(handle, inode, block, 1, 1);
                                error = -EIO;
                                goto cleanup;
                        }
index d0b7ca9..1852313 100644 (file)
@@ -20,6 +20,8 @@
 #include <linux/blkdev.h>
 #include <linux/magic.h>
 
+#include <linux/ext4_fs_i.h>
+
 /*
  * The second extended filesystem constants/structures
  */
 #define ext4_debug(f, a...)    do {} while (0)
 #endif
 
+#define EXT4_MULTIBLOCK_ALLOCATOR      1
+
+/* prefer goal again. length */
+#define EXT4_MB_HINT_MERGE             1
+/* blocks already reserved */
+#define EXT4_MB_HINT_RESERVED          2
+/* metadata is being allocated */
+#define EXT4_MB_HINT_METADATA          4
+/* first blocks in the file */
+#define EXT4_MB_HINT_FIRST             8
+/* search for the best chunk */
+#define EXT4_MB_HINT_BEST              16
+/* data is being allocated */
+#define EXT4_MB_HINT_DATA              32
+/* don't preallocate (for tails) */
+#define EXT4_MB_HINT_NOPREALLOC                64
+/* allocate for locality group */
+#define EXT4_MB_HINT_GROUP_ALLOC       128
+/* allocate goal blocks or none */
+#define EXT4_MB_HINT_GOAL_ONLY         256
+/* goal is meaningful */
+#define EXT4_MB_HINT_TRY_GOAL          512
+
+struct ext4_allocation_request {
+       /* target inode for block we're allocating */
+       struct inode *inode;
+       /* logical block in target inode */
+       ext4_lblk_t logical;
+       /* phys. target (a hint) */
+       ext4_fsblk_t goal;
+       /* the closest logical allocated block to the left */
+       ext4_lblk_t lleft;
+       /* phys. block for ^^^ */
+       ext4_fsblk_t pleft;
+       /* the closest logical allocated block to the right */
+       ext4_lblk_t lright;
+       /* phys. block for ^^^ */
+       ext4_fsblk_t pright;
+       /* how many blocks we want to allocate */
+       unsigned long len;
+       /* flags. see above EXT4_MB_HINT_* */
+       unsigned long flags;
+};
+
 /*
  * Special inodes numbers
  */
@@ -474,6 +520,7 @@ do {                                                                               \
 #define EXT4_MOUNT_JOURNAL_CHECKSUM    0x800000 /* Journal checksums */
 #define EXT4_MOUNT_JOURNAL_ASYNC_COMMIT        0x1000000 /* Journal Async Commit */
 #define EXT4_MOUNT_I_VERSION            0x2000000 /* i_version support */
+#define EXT4_MOUNT_MBALLOC             0x4000000 /* Buddy allocation support */
 /* Compatibility, for having both ext2_fs.h and ext4_fs.h included at once */
 #ifndef _LINUX_EXT2_FS_H
 #define clear_opt(o, opt)              o &= ~EXT4_MOUNT_##opt
@@ -912,7 +959,7 @@ extern ext4_fsblk_t ext4_new_blocks (handle_t *handle, struct inode *inode,
 extern ext4_fsblk_t ext4_new_blocks_old(handle_t *handle, struct inode *inode,
                        ext4_fsblk_t goal, unsigned long *count, int *errp);
 extern void ext4_free_blocks (handle_t *handle, struct inode *inode,
-                       ext4_fsblk_t block, unsigned long count);
+                       ext4_fsblk_t block, unsigned long count, int metadata);
 extern void ext4_free_blocks_sb (handle_t *handle, struct super_block *sb,
                                 ext4_fsblk_t block, unsigned long count,
                                unsigned long *pdquot_freed_blocks);
@@ -950,6 +997,20 @@ extern unsigned long ext4_count_dirs (struct super_block *);
 extern void ext4_check_inodes_bitmap (struct super_block *);
 extern unsigned long ext4_count_free (struct buffer_head *, unsigned);
 
+/* mballoc.c */
+extern long ext4_mb_stats;
+extern long ext4_mb_max_to_scan;
+extern int ext4_mb_init(struct super_block *, int);
+extern int ext4_mb_release(struct super_block *);
+extern ext4_fsblk_t ext4_mb_new_blocks(handle_t *,
+                               struct ext4_allocation_request *, int *);
+extern int ext4_mb_reserve_blocks(struct super_block *, int);
+extern void ext4_mb_discard_inode_preallocations(struct inode *);
+extern int __init init_ext4_mballoc(void);
+extern void exit_ext4_mballoc(void);
+extern void ext4_mb_free_blocks(handle_t *, struct inode *,
+               unsigned long, unsigned long, int, unsigned long *);
+
 
 /* inode.c */
 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
@@ -1080,6 +1141,19 @@ static inline void ext4_isize_set(struct ext4_inode *raw_inode, loff_t i_size)
        raw_inode->i_size_high = cpu_to_le32(i_size >> 32);
 }
 
+static inline
+struct ext4_group_info *ext4_get_group_info(struct super_block *sb,
+                                                       ext4_group_t group)
+{
+        struct ext4_group_info ***grp_info;
+        long indexv, indexh;
+        grp_info = EXT4_SB(sb)->s_group_info;
+        indexv = group >> (EXT4_DESC_PER_BLOCK_BITS(sb));
+        indexh = group & ((EXT4_DESC_PER_BLOCK(sb)) - 1);
+        return grp_info[indexv][indexh];
+}
+
+
 #define ext4_std_error(sb, errno)                              \
 do {                                                           \
        if ((errno))                                            \
index 4377d24..d5508d3 100644 (file)
@@ -158,6 +158,10 @@ struct ext4_inode_info {
         * struct timespec i_{a,c,m}time in the generic inode.
         */
        struct timespec i_crtime;
+
+       /* mballoc */
+       struct list_head i_prealloc_list;
+       spinlock_t i_prealloc_lock;
 };
 
 #endif /* _LINUX_EXT4_FS_I */
index 38a47ec..abaae2c 100644 (file)
@@ -91,6 +91,58 @@ struct ext4_sb_info {
        unsigned long s_ext_blocks;
        unsigned long s_ext_extents;
 #endif
+
+       /* for buddy allocator */
+       struct ext4_group_info ***s_group_info;
+       struct inode *s_buddy_cache;
+       long s_blocks_reserved;
+       spinlock_t s_reserve_lock;
+       struct list_head s_active_transaction;
+       struct list_head s_closed_transaction;
+       struct list_head s_committed_transaction;
+       spinlock_t s_md_lock;
+       tid_t s_last_transaction;
+       unsigned short *s_mb_offsets, *s_mb_maxs;
+
+       /* tunables */
+       unsigned long s_stripe;
+       unsigned long s_mb_stream_request;
+       unsigned long s_mb_max_to_scan;
+       unsigned long s_mb_min_to_scan;
+       unsigned long s_mb_stats;
+       unsigned long s_mb_order2_reqs;
+       unsigned long s_mb_group_prealloc;
+       /* where last allocation was done - for stream allocation */
+       unsigned long s_mb_last_group;
+       unsigned long s_mb_last_start;
+
+       /* history to debug policy */
+       struct ext4_mb_history *s_mb_history;
+       int s_mb_history_cur;
+       int s_mb_history_max;
+       int s_mb_history_num;
+       struct proc_dir_entry *s_mb_proc;
+       spinlock_t s_mb_history_lock;
+       int s_mb_history_filter;
+
+       /* stats for buddy allocator */
+       spinlock_t s_mb_pa_lock;
+       atomic_t s_bal_reqs;    /* number of reqs with len > 1 */
+       atomic_t s_bal_success; /* we found long enough chunks */
+       atomic_t s_bal_allocated;       /* in blocks */
+       atomic_t s_bal_ex_scanned;      /* total extents scanned */
+       atomic_t s_bal_goals;   /* goal hits */
+       atomic_t s_bal_breaks;  /* too long searches */
+       atomic_t s_bal_2orders; /* 2^order hits */
+       spinlock_t s_bal_lock;
+       unsigned long s_mb_buddies_generated;
+       unsigned long long s_mb_generation_time;
+       atomic_t s_mb_lost_chunks;
+       atomic_t s_mb_preallocated;
+       atomic_t s_mb_discarded;
+
+       /* locality groups */
+       struct ext4_locality_group *s_locality_groups;
 };
 
 #endif /* _LINUX_EXT4_FS_SB */