Whamcloud - gitweb
11f1ac07378ae467329bd6823694a4d64d4d321c
[fs/lustre-release.git] / ldiskfs / kernel_patches / patches / ext3-uninit-2.6-sles10.patch
1 Add support for the uninit_groups feature to the kernel.
2
3 Keep a high water mark of used inodes for each group to improve e2fsck time.
4 Block and inode bitmaps can be uninitialized on disk via a flag in the
5 group descriptor to avoid reading or scanning them at e2fsck time.
6 A checksum of each group descriptor is used to ensure that corruption in
7 the group descriptor's bit flags does not cause incorrect operation.
8
9 Index: linux-2.6.16.60-0.27/include/linux/ext3_fs.h
10 ===================================================================
11 --- linux-2.6.16.60-0.27.orig/include/linux/ext3_fs.h
12 +++ linux-2.6.16.60-0.27/include/linux/ext3_fs.h
13 @@ -153,16 +153,22 @@ struct ext3_allocation_request {
14   */
15  struct ext3_group_desc
16  {
17 -       __le32  bg_block_bitmap;                /* Blocks bitmap block */
18 -       __le32  bg_inode_bitmap;                /* Inodes bitmap block */
19 +       __le32  bg_block_bitmap;        /* Blocks bitmap block */
20 +       __le32  bg_inode_bitmap;        /* Inodes bitmap block */
21         __le32  bg_inode_table;         /* Inodes table block */
22         __le16  bg_free_blocks_count;   /* Free blocks count */
23         __le16  bg_free_inodes_count;   /* Free inodes count */
24         __le16  bg_used_dirs_count;     /* Directories count */
25 -       __u16   bg_pad;
26 -       __le32  bg_reserved[3];
27 +       __le16  bg_flags;               /* EXT3_BG_flags (UNINIT, etc) */
28 +       __le32  bg_reserved[2];         /* Likely block/inode bitmap checksum */
29 +       __le16  bg_itable_unused;       /* Unused inodes count */
30 +       __le16  bg_checksum;            /* crc16(sb_uuid+group+desc) */
31  };
32  
33 +#define EXT3_BG_INODE_UNINIT   0x0001 /* Inode table/bitmap not in use */
34 +#define EXT3_BG_BLOCK_UNINIT   0x0002 /* Block bitmap not in use */
35 +#define EXT3_BG_INODE_ZEROED   0x0004 /* On-disk itable initialized to zero */
36 +
37  /*
38   * Macro-instructions used to manage group descriptors
39   */
40 @@ -604,6 +610,7 @@ static inline int ext3_valid_inum(struct
41  #define EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER    0x0001
42  #define EXT3_FEATURE_RO_COMPAT_LARGE_FILE      0x0002
43  #define EXT3_FEATURE_RO_COMPAT_BTREE_DIR       0x0004
44 +#define EXT4_FEATURE_RO_COMPAT_GDT_CSUM                0x0010
45  #define EXT4_FEATURE_RO_COMPAT_DIR_NLINK       0x0020
46  
47  #define EXT3_FEATURE_INCOMPAT_COMPRESSION      0x0001
48 @@ -620,6 +627,7 @@ static inline int ext3_valid_inum(struct
49                                          EXT3_FEATURE_INCOMPAT_EXTENTS)
50  #define EXT3_FEATURE_RO_COMPAT_SUPP    (EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER| \
51                                          EXT3_FEATURE_RO_COMPAT_LARGE_FILE| \
52 +                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM| \
53                                          EXT4_FEATURE_RO_COMPAT_DIR_NLINK| \
54                                          EXT3_FEATURE_RO_COMPAT_BTREE_DIR)
55  
56 Index: linux-2.6.16.60-0.27/fs/ext3/resize.c
57 ===================================================================
58 --- linux-2.6.16.60-0.27.orig/fs/ext3/resize.c
59 +++ linux-2.6.16.60-0.27/fs/ext3/resize.c
60 @@ -19,6 +19,7 @@
61  #include <linux/errno.h>
62  #include <linux/slab.h>
63  
64 +#include "group.h"
65  
66  #define outside(b, first, last)        ((b) < (first) || (b) >= (last))
67  #define inside(b, first, last) ((b) >= (first) && (b) < (last))
68 @@ -134,25 +135,6 @@ static struct buffer_head *bclean(handle
69  }
70  
71  /*
72 - * To avoid calling the atomic setbit hundreds or thousands of times, we only
73 - * need to use it within a single byte (to ensure we get endianness right).
74 - * We can use memset for the rest of the bitmap as there are no other users.
75 - */
76 -static void mark_bitmap_end(int start_bit, int end_bit, char *bitmap)
77 -{
78 -       int i;
79 -
80 -       if (start_bit >= end_bit)
81 -               return;
82 -
83 -       ext3_debug("mark end bits +%d through +%d used\n", start_bit, end_bit);
84 -       for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++)
85 -               ext3_set_bit(i, bitmap);
86 -       if (i < end_bit)
87 -               memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3);
88 -}
89 -
90 -/*
91   * If we have fewer than thresh credits, extend by EXT3_MAX_TRANS_DATA.
92   * If that fails, restart the transaction & regain write access for the
93   * buffer head which is used for block_bitmap modifications.
94 @@ -865,6 +847,7 @@ int ext3_group_add(struct super_block *s
95         gdp->bg_inode_table = cpu_to_le32(input->inode_table);
96         gdp->bg_free_blocks_count = cpu_to_le16(input->free_blocks_count);
97         gdp->bg_free_inodes_count = cpu_to_le16(EXT3_INODES_PER_GROUP(sb));
98 +       gdp->bg_checksum = ext3_group_desc_csum(sbi, input->group, gdp);
99  
100         /*
101          * Make the new blocks and inodes valid next.  We do this before
102 Index: linux-2.6.16.60-0.27/fs/ext3/super.c
103 ===================================================================
104 --- linux-2.6.16.60-0.27.orig/fs/ext3/super.c
105 +++ linux-2.6.16.60-0.27/fs/ext3/super.c
106 @@ -42,6 +42,7 @@
107  #include "xattr.h"
108  #include "acl.h"
109  #include "namei.h"
110 +#include "group.h"
111  
112  static int ext3_load_journal(struct super_block *, struct ext3_super_block *,
113                              unsigned long journal_devnum);
114 @@ -1220,6 +1221,90 @@ static int ext3_setup_super(struct super
115         return res;
116  }
117  
118 +#if !defined(CONFIG_CRC16) && !defined(CONFIG_CRC16_MODULE)
119 +/** CRC table for the CRC-16. The poly is 0x8005 (x^16 + x^15 + x^2 + 1) */
120 +__u16 const crc16_table[256] = {
121 +       0x0000, 0xC0C1, 0xC181, 0x0140, 0xC301, 0x03C0, 0x0280, 0xC241,
122 +       0xC601, 0x06C0, 0x0780, 0xC741, 0x0500, 0xC5C1, 0xC481, 0x0440,
123 +       0xCC01, 0x0CC0, 0x0D80, 0xCD41, 0x0F00, 0xCFC1, 0xCE81, 0x0E40,
124 +       0x0A00, 0xCAC1, 0xCB81, 0x0B40, 0xC901, 0x09C0, 0x0880, 0xC841,
125 +       0xD801, 0x18C0, 0x1980, 0xD941, 0x1B00, 0xDBC1, 0xDA81, 0x1A40,
126 +       0x1E00, 0xDEC1, 0xDF81, 0x1F40, 0xDD01, 0x1DC0, 0x1C80, 0xDC41,
127 +       0x1400, 0xD4C1, 0xD581, 0x1540, 0xD701, 0x17C0, 0x1680, 0xD641,
128 +       0xD201, 0x12C0, 0x1380, 0xD341, 0x1100, 0xD1C1, 0xD081, 0x1040,
129 +       0xF001, 0x30C0, 0x3180, 0xF141, 0x3300, 0xF3C1, 0xF281, 0x3240,
130 +       0x3600, 0xF6C1, 0xF781, 0x3740, 0xF501, 0x35C0, 0x3480, 0xF441,
131 +       0x3C00, 0xFCC1, 0xFD81, 0x3D40, 0xFF01, 0x3FC0, 0x3E80, 0xFE41,
132 +       0xFA01, 0x3AC0, 0x3B80, 0xFB41, 0x3900, 0xF9C1, 0xF881, 0x3840,
133 +       0x2800, 0xE8C1, 0xE981, 0x2940, 0xEB01, 0x2BC0, 0x2A80, 0xEA41,
134 +       0xEE01, 0x2EC0, 0x2F80, 0xEF41, 0x2D00, 0xEDC1, 0xEC81, 0x2C40,
135 +       0xE401, 0x24C0, 0x2580, 0xE541, 0x2700, 0xE7C1, 0xE681, 0x2640,
136 +       0x2200, 0xE2C1, 0xE381, 0x2340, 0xE101, 0x21C0, 0x2080, 0xE041,
137 +       0xA001, 0x60C0, 0x6180, 0xA141, 0x6300, 0xA3C1, 0xA281, 0x6240,
138 +       0x6600, 0xA6C1, 0xA781, 0x6740, 0xA501, 0x65C0, 0x6480, 0xA441,
139 +       0x6C00, 0xACC1, 0xAD81, 0x6D40, 0xAF01, 0x6FC0, 0x6E80, 0xAE41,
140 +       0xAA01, 0x6AC0, 0x6B80, 0xAB41, 0x6900, 0xA9C1, 0xA881, 0x6840,
141 +       0x7800, 0xB8C1, 0xB981, 0x7940, 0xBB01, 0x7BC0, 0x7A80, 0xBA41,
142 +       0xBE01, 0x7EC0, 0x7F80, 0xBF41, 0x7D00, 0xBDC1, 0xBC81, 0x7C40,
143 +       0xB401, 0x74C0, 0x7580, 0xB541, 0x7700, 0xB7C1, 0xB681, 0x7640,
144 +       0x7200, 0xB2C1, 0xB381, 0x7340, 0xB101, 0x71C0, 0x7080, 0xB041,
145 +       0x5000, 0x90C1, 0x9181, 0x5140, 0x9301, 0x53C0, 0x5280, 0x9241,
146 +       0x9601, 0x56C0, 0x5780, 0x9741, 0x5500, 0x95C1, 0x9481, 0x5440,
147 +       0x9C01, 0x5CC0, 0x5D80, 0x9D41, 0x5F00, 0x9FC1, 0x9E81, 0x5E40,
148 +       0x5A00, 0x9AC1, 0x9B81, 0x5B40, 0x9901, 0x59C0, 0x5880, 0x9841,
149 +       0x8801, 0x48C0, 0x4980, 0x8941, 0x4B00, 0x8BC1, 0x8A81, 0x4A40,
150 +       0x4E00, 0x8EC1, 0x8F81, 0x4F40, 0x8D01, 0x4DC0, 0x4C80, 0x8C41,
151 +       0x4400, 0x84C1, 0x8581, 0x4540, 0x8701, 0x47C0, 0x4680, 0x8641,
152 +       0x8201, 0x42C0, 0x4380, 0x8341, 0x4100, 0x81C1, 0x8081, 0x4040
153 +};
154 +
155 +static inline __u16 crc16_byte(__u16 crc, const __u8 data)
156 +{
157 +       return (crc >> 8) ^ crc16_table[(crc ^ data) & 0xff];
158 +}
159 +
160 +__u16 crc16(__u16 crc, __u8 const *buffer, size_t len)
161 +{
162 +       while (len--)
163 +               crc = crc16_byte(crc, *buffer++);
164 +       return crc;
165 +}
166 +#endif
167 +
168 +__le16 ext3_group_desc_csum(struct ext3_sb_info *sbi, __u32 block_group,
169 +                           struct ext3_group_desc *gdp)
170 +{
171 +       __u16 crc = 0;
172 +
173 +       if (sbi->s_es->s_feature_ro_compat &
174 +           cpu_to_le32(EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) {
175 +               int offset = offsetof(struct ext3_group_desc, bg_checksum);
176 +               __le32 le_group = cpu_to_le32(block_group);
177 +
178 +               crc = crc16(~0, sbi->s_es->s_uuid, sizeof(sbi->s_es->s_uuid));
179 +               crc = crc16(crc, (__u8 *)&le_group, sizeof(le_group));
180 +               crc = crc16(crc, (__u8 *)gdp, offset);
181 +               offset += sizeof(gdp->bg_checksum); /* skip checksum */
182 +               BUG_ON(offset != sizeof(*gdp)); /* XXX handle s_desc_size */
183 +               /* for checksum of struct ext4_group_desc do the rest...
184 +               if (offset < sbi->s_es->s_desc_size) {
185 +                       crc = crc16(crc, (__u8 *)gdp + offset,
186 +                                   sbi->s_es->s_desc_size - offset);
187 +                */
188 +       }
189 +
190 +       return cpu_to_le16(crc);
191 +}
192 +
193 +int ext3_group_desc_csum_verify(struct ext3_sb_info *sbi, __u32 block_group,
194 +                               struct ext3_group_desc *gdp)
195 +{
196 +       if (gdp->bg_checksum != ext3_group_desc_csum(sbi, block_group, gdp))
197 +               return 0;
198 +
199 +       return 1;
200 +}
201 +
202  /* Called at mount-time, super-block is locked */
203  static int ext3_check_descriptors (struct super_block * sb)
204  {
205 @@ -1269,6 +1354,13 @@ static int ext3_check_descriptors (struc
206                                         le32_to_cpu(gdp->bg_inode_table));
207                         return 0;
208                 }
209 +               if (!ext3_group_desc_csum_verify(sbi, i, gdp)) {
210 +                       ext3_error(sb, __FUNCTION__,
211 +                                  "Checksum for group %d failed (%u!=%u)\n", i,
212 +                                  le16_to_cpu(ext3_group_desc_csum(sbi,i,gdp)),
213 +                                  le16_to_cpu(gdp->bg_checksum));
214 +                       return 0;
215 +               }
216                 block += EXT3_BLOCKS_PER_GROUP(sb);
217                 gdp++;
218         }
219 Index: linux-2.6.16.60-0.27/fs/ext3/group.h
220 ===================================================================
221 --- /dev/null
222 +++ linux-2.6.16.60-0.27/fs/ext3/group.h
223 @@ -0,0 +1,30 @@
224 +/*
225 + *  linux/fs/ext3/group.h
226 + *
227 + * Copyright 2008 Sun Microsystems, Inc
228 + *
229 + * Author: Andreas Dilger <adilger@clusterfs.com>
230 + */
231 +
232 +#ifndef _LINUX_EXT3_GROUP_H
233 +#define _LINUX_EXT3_GROUP_H
234 +#if defined(CONFIG_CRC16) || defined(CONFIG_CRC16_MODULE)
235 +#include <linux/crc16.h>
236 +#endif
237 +
238 +extern __le16 ext3_group_desc_csum(struct ext3_sb_info *sbi, __u32 group,
239 +                                  struct ext3_group_desc *gdp);
240 +extern int ext3_group_desc_csum_verify(struct ext3_sb_info *sbi, __u32 group,
241 +                                      struct ext3_group_desc *gdp);
242 +struct buffer_head *read_block_bitmap(struct super_block *sb,
243 +                                     unsigned int block_group);
244 +extern unsigned ext3_init_block_bitmap(struct super_block *sb,
245 +                                      struct buffer_head *bh, int group,
246 +                                      struct ext3_group_desc *desc);
247 +#define ext3_free_blocks_after_init(sb, group, desc)                   \
248 +               ext3_init_block_bitmap(sb, NULL, group, desc)
249 +extern unsigned ext3_init_inode_bitmap(struct super_block *sb,
250 +                                      struct buffer_head *bh, int group,
251 +                                      struct ext3_group_desc *desc);
252 +extern void mark_bitmap_end(int start_bit, int end_bit, char *bitmap);
253 +#endif /* _LINUX_EXT3_GROUP_H */
254 Index: linux-2.6.16.60-0.27/fs/ext3/ialloc.c
255 ===================================================================
256 --- linux-2.6.16.60-0.27.orig/fs/ext3/ialloc.c
257 +++ linux-2.6.16.60-0.27/fs/ext3/ialloc.c
258 @@ -28,6 +28,7 @@
259  
260  #include "xattr.h"
261  #include "acl.h"
262 +#include "group.h"
263  
264  /*
265   * ialloc.c contains the inodes allocation and deallocation routines
266 @@ -43,6 +44,52 @@
267   * the free blocks count in the block.
268   */
269  
270 +/*
271 + * To avoid calling the atomic setbit hundreds or thousands of times, we only
272 + * need to use it within a single byte (to ensure we get endianness right).
273 + * We can use memset for the rest of the bitmap as there are no other users.
274 + */
275 +void mark_bitmap_end(int start_bit, int end_bit, char *bitmap)
276 +{
277 +       int i;
278 +
279 +       if (start_bit >= end_bit)
280 +               return;
281 +
282 +       ext3_debug("mark end bits +%d through +%d used\n", start_bit, end_bit);
283 +       for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++)
284 +               ext3_set_bit(i, bitmap);
285 +       if (i < end_bit)
286 +               memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3);
287 +}
288 +
289 +/* Initializes an uninitialized inode bitmap */
290 +unsigned ext3_init_inode_bitmap(struct super_block *sb,
291 +                               struct buffer_head *bh, int block_group,
292 +                               struct ext3_group_desc *gdp)
293 +{
294 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
295 +
296 +       J_ASSERT_BH(bh, buffer_locked(bh));
297 +
298 +       /* If checksum is bad mark all blocks and inodes use to prevent
299 +        * allocation, essentially implementing a per-group read-only flag. */
300 +       if (!ext3_group_desc_csum_verify(sbi, block_group, gdp)) {
301 +               ext3_error(sb, __FUNCTION__, "Checksum bad for group %u\n",
302 +                          block_group);
303 +               gdp->bg_free_blocks_count = 0;
304 +               gdp->bg_free_inodes_count = 0;
305 +               gdp->bg_itable_unused = 0;
306 +               memset(bh->b_data, 0xff, sb->s_blocksize);
307 +               return 0;
308 +       }
309 +
310 +       memset(bh->b_data, 0, (EXT3_INODES_PER_GROUP(sb) + 7) / 8);
311 +       mark_bitmap_end(EXT3_INODES_PER_GROUP(sb), EXT3_BLOCKS_PER_GROUP(sb),
312 +                       bh->b_data);
313 +
314 +       return EXT3_INODES_PER_GROUP(sb);
315 +}
316  
317  /*
318   * Read the inode allocation bitmap for a given block_group, reading
319 @@ -59,8 +106,19 @@ read_inode_bitmap(struct super_block * s
320         desc = ext3_get_group_desc(sb, block_group, NULL);
321         if (!desc)
322                 goto error_out;
323 -
324 -       bh = sb_bread(sb, le32_to_cpu(desc->bg_inode_bitmap));
325 +       if (desc->bg_flags & cpu_to_le16(EXT3_BG_INODE_UNINIT)) {
326 +               bh = sb_getblk(sb, le32_to_cpu(desc->bg_inode_bitmap));
327 +               if (!buffer_uptodate(bh)) {
328 +                       lock_buffer(bh);
329 +                       if (!buffer_uptodate(bh)) {
330 +                               ext3_init_inode_bitmap(sb, bh,block_group,desc);
331 +                               set_buffer_uptodate(bh);
332 +                       }
333 +                       unlock_buffer(bh);
334 +               }
335 +       } else {
336 +               bh = sb_bread(sb, le32_to_cpu(desc->bg_inode_bitmap));
337 +       }
338         if (!bh)
339                 ext3_error(sb, "read_inode_bitmap",
340                             "Cannot read inode bitmap - "
341 @@ -169,6 +227,8 @@ void ext3_free_inode (handle_t *handle, 
342                         if (is_directory)
343                                 gdp->bg_used_dirs_count = cpu_to_le16(
344                                   le16_to_cpu(gdp->bg_used_dirs_count) - 1);
345 +                       gdp->bg_checksum = ext3_group_desc_csum(sbi,block_group,
346 +                                                               gdp);
347                         spin_unlock(sb_bgl_lock(sbi, block_group));
348                         percpu_counter_inc(&sbi->s_freeinodes_counter);
349                         if (is_directory)
350 @@ -452,7 +512,7 @@ struct inode *ext3_new_inode(handle_t *h
351         struct ext3_sb_info *sbi;
352         int err = 0;
353         struct inode *ret;
354 -       int i;
355 +       int i, free = 0;
356  
357         /* Cannot create files in a deleted directory */
358         if (!dir || !dir->i_nlink)
359 @@ -569,11 +629,13 @@ repeat_in_this_group:
360         goto out;
361  
362  got:
363 -       ino += group * EXT3_INODES_PER_GROUP(sb) + 1;
364 -       if (ino < EXT3_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) {
365 -               ext3_error (sb, "ext3_new_inode",
366 -                           "reserved inode or inode > inodes count - "
367 -                           "block_group = %d, inode=%lu", group, ino);
368 +       ino++;
369 +       if ((group == 0 && ino < EXT3_FIRST_INO(sb)) ||
370 +           ino > EXT3_INODES_PER_GROUP(sb)) {
371 +               ext3_error(sb, __FUNCTION__,
372 +                          "reserved inode or inode > inodes count - "
373 +                          "block_group = %d, inode=%lu", group,
374 +                          ino + group * EXT3_INODES_PER_GROUP(sb));
375                 err = -EIO;
376                 goto fail;
377         }
378 @@ -581,13 +643,65 @@ got:
379         BUFFER_TRACE(bh2, "get_write_access");
380         err = ext3_journal_get_write_access(handle, bh2);
381         if (err) goto fail;
382 +
383 +       /* We may have to initialize the block bitmap if it isn't already */
384 +       if (EXT3_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) &&
385 +           gdp->bg_flags & cpu_to_le16(EXT3_BG_BLOCK_UNINIT)) {
386 +               struct buffer_head *block_bh = read_block_bitmap(sb, group);
387 +
388 +               BUFFER_TRACE(block_bh, "get block bitmap access");
389 +               err = ext3_journal_get_write_access(handle, block_bh);
390 +               if (err) {
391 +                       brelse(block_bh);
392 +                       goto fail;
393 +               }
394 +
395 +               free = 0;
396 +               spin_lock(sb_bgl_lock(sbi, group));
397 +               /* recheck and clear flag under lock if we still need to */
398 +               if (gdp->bg_flags & cpu_to_le16(EXT3_BG_BLOCK_UNINIT)) {
399 +                       gdp->bg_flags &= cpu_to_le16(~EXT3_BG_BLOCK_UNINIT);
400 +                       free = ext3_free_blocks_after_init(sb, group, gdp);
401 +                       gdp->bg_free_blocks_count = cpu_to_le16(free);
402 +               }
403 +               spin_unlock(sb_bgl_lock(sbi, group));
404 +
405 +               /* Don't need to dirty bitmap block if we didn't change it */
406 +               if (free) {
407 +                       BUFFER_TRACE(block_bh, "dirty block bitmap");
408 +                       err = ext3_journal_dirty_metadata(handle, block_bh);
409 +               }
410 +
411 +               brelse(block_bh);
412 +               if (err)
413 +                       goto fail;
414 +       }
415 +
416         spin_lock(sb_bgl_lock(sbi, group));
417 +       /* If we didn't allocate from within the initialized part of the inode
418 +        * table then we need to initialize up to this inode. */
419 +       if (EXT3_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) {
420 +               if (gdp->bg_flags & cpu_to_le16(EXT3_BG_INODE_UNINIT)) {
421 +                       gdp->bg_flags &= cpu_to_le16(~EXT3_BG_INODE_UNINIT);
422 +                       free = 0;
423 +               } else {
424 +                       free = EXT3_INODES_PER_GROUP(sb) -
425 +                               le16_to_cpu(gdp->bg_itable_unused);
426 +               }
427 +
428 +               if (ino > free) {
429 +                       gdp->bg_itable_unused =
430 +                               cpu_to_le16(EXT3_INODES_PER_GROUP(sb) - ino);
431 +               }
432 +       }
433 +
434         gdp->bg_free_inodes_count =
435                 cpu_to_le16(le16_to_cpu(gdp->bg_free_inodes_count) - 1);
436         if (S_ISDIR(mode)) {
437                 gdp->bg_used_dirs_count =
438                         cpu_to_le16(le16_to_cpu(gdp->bg_used_dirs_count) + 1);
439         }
440 +       gdp->bg_checksum = ext3_group_desc_csum(sbi, group, gdp);
441         spin_unlock(sb_bgl_lock(sbi, group));
442         BUFFER_TRACE(bh2, "call ext3_journal_dirty_metadata");
443         err = ext3_journal_dirty_metadata(handle, bh2);
444 @@ -609,7 +723,7 @@ got:
445                 inode->i_gid = current->fsgid;
446         inode->i_mode = mode;
447  
448 -       inode->i_ino = ino;
449 +       inode->i_ino = ino + group * EXT3_INODES_PER_GROUP(sb);
450         /* This is the optimal IO size (for stat), not the fs block size */
451         inode->i_blksize = PAGE_SIZE;
452         inode->i_blocks = 0;
453 Index: linux-2.6.16.60-0.27/fs/ext3/mballoc.c
454 ===================================================================
455 --- linux-2.6.16.60-0.27.orig/fs/ext3/mballoc.c
456 +++ linux-2.6.16.60-0.27/fs/ext3/mballoc.c
457 @@ -36,6 +36,8 @@
458  #include <linux/seq_file.h>
459  #include <linux/version.h>
460  
461 +#include "group.h"
462 +
463  /*
464   * MUSTDO:
465   *   - test ext3_ext_search_left() and ext3_ext_search_right()
466 @@ -323,6 +325,7 @@ struct ext3_group_info {
467         unsigned long   bb_state;
468         unsigned long   bb_tid;
469         struct ext3_free_metadata *bb_md_cur;
470 +       struct ext3_group_desc *bb_gdp;
471         unsigned short  bb_first_free;
472         unsigned short  bb_free;
473         unsigned short  bb_fragments;
474 @@ -943,10 +946,7 @@ static int ext3_mb_init_cache(struct pag
475                 if (first_group + i >= EXT3_SB(sb)->s_groups_count)
476                         break;
477  
478 -               err = -EIO;
479 -               desc = ext3_get_group_desc(sb, first_group + i, NULL);
480 -               if (desc == NULL)
481 -                       goto out;
482 +               desc = EXT3_GROUP_INFO(sb, first_group + i)->bb_gdp;
483  
484                 err = -ENOMEM;
485                 bh[i] = sb_getblk(sb, le32_to_cpu(desc->bg_block_bitmap));
486 @@ -961,7 +961,12 @@ static int ext3_mb_init_cache(struct pag
487                         unlock_buffer(bh[i]);
488                         continue;
489                 }
490 -
491 +               if (desc->bg_flags & cpu_to_le16(EXT3_BG_BLOCK_UNINIT)) {
492 +                       ext3_init_block_bitmap(sb, bh[i], first_group + i,desc);
493 +                       set_buffer_uptodate(bh[i]);
494 +                       unlock_buffer(bh[i]);
495 +                       continue;
496 +               }
497                 get_bh(bh[i]);
498                 bh[i]->b_end_io = end_buffer_read_sync;
499                 submit_bh(READ, bh[i]);
500 @@ -1733,6 +1738,10 @@ static int ext3_mb_good_group(struct ext
501         switch (cr) {
502                 case 0:
503                         BUG_ON(ac->ac_2order == 0);
504 +                       /* If this group is uninitialized, skip it initially */
505 +                       if (grp->bb_gdp->bg_flags &
506 +                           cpu_to_le16(EXT3_BG_BLOCK_UNINIT))
507 +                               return 0;
508                         bits = ac->ac_sb->s_blocksize_bits + 1;
509                         for (i = ac->ac_2order; i <= bits; i++)
510                                 if (grp->bb_counters[i] > 0)
511 @@ -1826,7 +1835,9 @@ repeat:
512                         }
513  
514                         ac->ac_groups_scanned++;
515 -                       if (cr == 0)
516 +                       if (cr == 0 || (e3b.bd_info->bb_gdp->bg_flags &
517 +                                       cpu_to_le16(EXT3_BG_BLOCK_UNINIT) &&
518 +                                       ac->ac_2order != 0))
519                                 ext3_mb_simple_scan_group(ac, &e3b);
520                         else if (cr == 1 && ac->ac_g_ex.fe_len == sbi->s_stripe)
521                                 ext3_mb_scan_aligned(ac, &e3b);
522 @@ -2306,12 +2317,13 @@ int ext3_mb_init_backend(struct super_bl
523                         i--;
524                         goto err_freebuddy;
525                 }
526 +               memset(meta_group_info[j], 0, len);
527                 desc = ext3_get_group_desc(sb, i, NULL);
528 +               meta_group_info[j]->bb_gdp = desc;
529                 if (desc == NULL) {
530                         printk(KERN_ERR"EXT3-fs: can't read descriptor %u\n",i);
531                         goto err_freebuddy;
532                 }
533 -               memset(meta_group_info[j], 0, len);
534                 set_bit(EXT3_GROUP_INFO_NEED_INIT_BIT,
535                         &meta_group_info[j]->bb_state);
536  
537 @@ -2945,9 +2957,17 @@ int ext3_mb_mark_diskspace_used(struct e
538                     ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len);
539  
540         spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
541 +       if (gdp->bg_flags & cpu_to_le16(EXT3_BG_BLOCK_UNINIT)) {
542 +               gdp->bg_flags &= cpu_to_le16(~EXT3_BG_BLOCK_UNINIT);
543 +               gdp->bg_free_blocks_count =
544 +                       cpu_to_le16(ext3_free_blocks_after_init(sb,
545 +                                                           ac->ac_b_ex.fe_group,
546 +                                                           gdp));
547 +       }
548         gdp->bg_free_blocks_count =
549                 cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count)
550                                 - ac->ac_b_ex.fe_len);
551 +       gdp->bg_checksum = ext3_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp);
552         spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group));
553         percpu_counter_mod(&sbi->s_freeblocks_counter, - ac->ac_b_ex.fe_len);
554  
555 @@ -4357,6 +4377,7 @@ do_more:
556         spin_lock(sb_bgl_lock(sbi, block_group));
557         gdp->bg_free_blocks_count =
558                 cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) + count);
559 +       gdp->bg_checksum = ext3_group_desc_csum(sbi, block_group, gdp);
560         spin_unlock(sb_bgl_lock(sbi, block_group));
561         percpu_counter_mod(&sbi->s_freeblocks_counter, count);
562  
563 Index: linux-2.6.16.60-0.27/fs/ext3/balloc.c
564 ===================================================================
565 --- linux-2.6.16.60-0.27.orig/fs/ext3/balloc.c
566 +++ linux-2.6.16.60-0.27/fs/ext3/balloc.c
567 @@ -21,6 +21,7 @@
568  #include <linux/quotaops.h>
569  #include <linux/buffer_head.h>
570  
571 +#include "group.h"
572  /*
573   * balloc.c contains the blocks allocation and deallocation routines
574   */
575 @@ -74,6 +75,83 @@ struct ext3_group_desc * ext3_get_group_
576         return desc + offset;
577  }
578  
579 +/* Initializes an uninitialized block bitmap if given, and returns the
580 + * number of blocks free in the group. */
581 +unsigned ext3_init_block_bitmap(struct super_block *sb, struct buffer_head *bh,
582 +                               int block_group, struct ext3_group_desc *gdp)
583 +{
584 +       unsigned long start;
585 +       int bit, bit_max;
586 +       unsigned free_blocks;
587 +       struct ext3_sb_info *sbi = EXT3_SB(sb);
588 +
589 +       if (bh) {
590 +               J_ASSERT_BH(bh, buffer_locked(bh));
591 +
592 +               /* If checksum is bad mark all blocks use to prevent allocation,
593 +                * essentially implementing a per-group read-only flag. */
594 +               if (!ext3_group_desc_csum_verify(sbi, block_group, gdp)) {
595 +                       ext3_error(sb, __FUNCTION__,
596 +                                  "Checksum bad for group %u\n", block_group);
597 +                       gdp->bg_free_blocks_count = 0;
598 +                       gdp->bg_free_inodes_count = 0;
599 +                       gdp->bg_itable_unused = 0;
600 +                       memset(bh->b_data, 0xff, sb->s_blocksize);
601 +                       return 0;
602 +               }
603 +               memset(bh->b_data, 0, sb->s_blocksize);
604 +       }
605 +
606 +       /* Check for superblock and gdt backups in this group */
607 +       bit_max = ext3_bg_has_super(sb, block_group);
608 +
609 +       if (!EXT3_HAS_INCOMPAT_FEATURE(sb, EXT3_FEATURE_INCOMPAT_META_BG) ||
610 +           block_group < le32_to_cpu(sbi->s_es->s_first_meta_bg) *
611 +                         sbi->s_desc_per_block) {
612 +               if (bit_max) {
613 +                       bit_max += ext3_bg_num_gdb(sb, block_group);
614 +                       bit_max +=le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks);
615 +               }
616 +       } else { /* For META_BG_BLOCK_GROUPS */
617 +               int group_rel = (block_group -
618 +                                le32_to_cpu(sbi->s_es->s_first_meta_bg)) %
619 +                               EXT3_DESC_PER_BLOCK(sb);
620 +               if (group_rel == 0 || group_rel == 1 ||
621 +                   (group_rel == EXT3_DESC_PER_BLOCK(sb) - 1))
622 +                       bit_max += 1;
623 +       }
624 +
625 +       /* Last and first groups are always initialized */
626 +       free_blocks = EXT3_BLOCKS_PER_GROUP(sb) - bit_max;
627 +
628 +       if (bh) {
629 +               for (bit = 0; bit < bit_max; bit++)
630 +                       ext3_set_bit(bit, bh->b_data);
631 +
632 +               start = block_group * EXT3_BLOCKS_PER_GROUP(sb) +
633 +                       le32_to_cpu(sbi->s_es->s_first_data_block);
634 +
635 +               /* Set bits for block and inode bitmaps, and inode table */
636 +               ext3_set_bit(le32_to_cpu(gdp->bg_block_bitmap) - start,
637 +                            bh->b_data);
638 +               ext3_set_bit(le32_to_cpu(gdp->bg_inode_bitmap) - start,
639 +                            bh->b_data);
640 +               for (bit = le32_to_cpu(gdp->bg_inode_table) - start,
641 +                    bit_max = bit + sbi->s_itb_per_group; bit < bit_max; bit++)
642 +                       ext3_set_bit(bit, bh->b_data);
643 +
644 +               /*
645 +                * Also if the number of blocks within the group is
646 +                * less than the blocksize * 8 ( which is the size
647 +                * of bitmap ), set rest of the block bitmap to 1
648 +                */
649 +               mark_bitmap_end(EXT3_BLOCKS_PER_GROUP(sb), sb->s_blocksize * 8,
650 +                               bh->b_data);
651 +       }
652 +
653 +       return free_blocks - sbi->s_itb_per_group - 2;
654 +}
655 +
656  /*
657   * Read the bitmap for a given block_group, reading into the specified 
658   * slot in the superblock's bitmap cache.
659 @@ -89,7 +167,19 @@ read_block_bitmap(struct super_block *sb
660         desc = ext3_get_group_desc (sb, block_group, NULL);
661         if (!desc)
662                 goto error_out;
663 -       bh = sb_bread(sb, le32_to_cpu(desc->bg_block_bitmap));
664 +       if (desc->bg_flags & cpu_to_le16(EXT3_BG_BLOCK_UNINIT)) {
665 +               bh = sb_getblk(sb, le32_to_cpu(desc->bg_block_bitmap));
666 +               if (!buffer_uptodate(bh)) {
667 +                       lock_buffer(bh);
668 +                       if (!buffer_uptodate(bh)) {
669 +                               ext3_init_block_bitmap(sb, bh,block_group,desc);
670 +                               set_buffer_uptodate(bh);
671 +                       }
672 +                       unlock_buffer(bh);
673 +               }
674 +       } else {
675 +               bh = sb_bread(sb, le32_to_cpu(desc->bg_block_bitmap));
676 +       }
677         if (!bh)
678                 ext3_error (sb, "read_block_bitmap",
679                             "Cannot read block bitmap - "
680 @@ -468,6 +558,7 @@ do_more:
681         desc->bg_free_blocks_count =
682                 cpu_to_le16(le16_to_cpu(desc->bg_free_blocks_count) +
683                         group_freed);
684 +       desc->bg_checksum = ext3_group_desc_csum(sbi, block_group, desc);
685         spin_unlock(sb_bgl_lock(sbi, block_group));
686         percpu_counter_mod(&sbi->s_freeblocks_counter, count);
687  
688 @@ -1378,8 +1469,11 @@ allocated:
689                         ret_block, goal_hits, goal_attempts);
690  
691         spin_lock(sb_bgl_lock(sbi, group_no));
692 +       if (gdp->bg_flags & cpu_to_le16(EXT3_BG_BLOCK_UNINIT))
693 +               gdp->bg_flags &= cpu_to_le16(~EXT3_BG_BLOCK_UNINIT);
694         gdp->bg_free_blocks_count =
695                         cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) - 1);
696 +       gdp->bg_checksum = ext3_group_desc_csum(sbi, group_no, gdp);
697         spin_unlock(sb_bgl_lock(sbi, group_no));
698         percpu_counter_mod(&sbi->s_freeblocks_counter, -1);
699