ldiskfs/kernel_patches/patches/sles11sp2/ext4-mballoc-extra-checks.patch

   1 ---
   2  fs/ext4/ext4.h    |    1
   3  fs/ext4/mballoc.c |  108 ++++++++++++++++++++++++++++++++++++++++++++++--------
   4  fs/ext4/mballoc.h |    2 -
   5  3 files changed, 95 insertions(+), 16 deletions(-)
   6
   7 --- a/fs/ext4/ext4.h
   8 +++ b/fs/ext4/ext4.h
   9 @@ -2098,6 +2098,7 @@ struct ext4_group_info {
  10         ext4_grpblk_t   bb_fragments;   /* nr of freespace fragments */
  11         ext4_grpblk_t   bb_largest_free_order;/* order of largest frag in BG */
  12         struct          list_head bb_prealloc_list;
  13 +       unsigned long   bb_prealloc_nr;
  14  #ifdef DOUBLE_CHECK
  15         void            *bb_bitmap;
  16  #endif
  17 --- a/fs/ext4/mballoc.c
  18 +++ b/fs/ext4/mballoc.c
  19 @@ -352,7 +352,7 @@ static const char *ext4_groupinfo_slab_n
  20         "ext4_groupinfo_64k", "ext4_groupinfo_128k"
  21  };
  22
  23 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  24 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  25                                         ext4_group_t group);
  26  static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
  27                                                 ext4_group_t group);
  28 @@ -702,7 +702,7 @@ mb_set_largest_free_order(struct super_b
  29  }
  30
  31  static noinline_for_stack
  32 -void ext4_mb_generate_buddy(struct super_block *sb,
  33 +int ext4_mb_generate_buddy(struct super_block *sb,
  34                                 void *buddy, void *bitmap, ext4_group_t group)
  35  {
  36         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  37 @@ -734,14 +734,19 @@ void ext4_mb_generate_buddy(struct super
  38         grp->bb_fragments = fragments;
  39
  40         if (free != grp->bb_free) {
  41 +               struct ext4_group_desc *gdp;
  42 +               gdp = ext4_get_group_desc (sb, group, NULL);
  43                 ext4_grp_locked_error(sb, group, 0, 0,
  44 -                                     "%u blocks in bitmap, %u in gd",
  45 -                                     free, grp->bb_free);
  46 +                                     "%u blocks in bitmap, %u in bb, %u in gd",
  47 +                                     free, grp->bb_free,
  48 +                                     ext4_free_blks_count(sb, gdp));
  49 +
  50                 /*
  51                  * If we intent to continue, we consider group descritor
  52                  * corrupt and update bb_free using bitmap value
  53                  */
  54                 grp->bb_free = free;
  55 +               return -EIO;
  56         }
  57         mb_set_largest_free_order(sb, grp);
  58
  59 @@ -752,6 +757,8 @@ void ext4_mb_generate_buddy(struct super
  60         EXT4_SB(sb)->s_mb_buddies_generated++;
  61         EXT4_SB(sb)->s_mb_generation_time += period;
  62         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
  63 +
  64 +       return 0;
  65  }
  66
  67  /* The buddy information is attached the buddy cache inode
  68 @@ -898,7 +905,7 @@ static int ext4_mb_init_cache(struct pag
  69
  70         err = 0;
  71         first_block = page->index * blocks_per_page;
  72 -       for (i = 0; i < blocks_per_page; i++) {
  73 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  74                 int group;
  75
  76                 group = (first_block + i) >> 1;
  77 @@ -939,7 +946,7 @@ static int ext4_mb_init_cache(struct pag
  78                         ext4_lock_group(sb, group);
  79                         /* init the buddy */
  80                         memset(data, 0xff, blocksize);
  81 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  82 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  83                         ext4_unlock_group(sb, group);
  84                         incore = NULL;
  85                 } else {
  86 @@ -954,7 +961,7 @@ static int ext4_mb_init_cache(struct pag
  87                         memcpy(data, bitmap, blocksize);
  88
  89                         /* mark all preallocated blks used in in-core bitmap */
  90 -                       ext4_mb_generate_from_pa(sb, data, group);
  91 +                       err = ext4_mb_generate_from_pa(sb, data, group);
  92                         ext4_mb_generate_from_freelist(sb, data, group);
  93                         ext4_unlock_group(sb, group);
  94
  95 @@ -964,7 +971,8 @@ static int ext4_mb_init_cache(struct pag
  96                         incore = data;
  97                 }
  98         }
  99 -       SetPageUptodate(page);
 100 +       if (likely(err == 0))
 101 +               SetPageUptodate(page);
 102
 103  out:
 104         if (bh) {
 105 @@ -2143,9 +2151,11 @@ static void *ext4_mb_seq_groups_next(str
 106  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 107  {
 108         struct super_block *sb = seq->private;
 109 +       struct ext4_group_desc *gdp;
 110         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 111         int i;
 112         int err;
 113 +       int free = 0;
 114         struct ext4_buddy e4b;
 115         struct sg {
 116                 struct ext4_group_info info;
 117 @@ -2154,10 +2164,10 @@ static int ext4_mb_seq_groups_show(struc
 118
 119         group--;
 120         if (group == 0)
 121 -               seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
 122 +               seq_printf(seq, "#%-5s: %-5s %-5s %-5s %-5s %-5s"
 123                                 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
 124                                   "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
 125 -                          "group", "free", "frags", "first",
 126 +                          "group", "free", "free", "frags", "first", "pa",
 127                            "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
 128                            "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
 129
 130 @@ -2168,13 +2178,20 @@ static int ext4_mb_seq_groups_show(struc
 131                 seq_printf(seq, "#%-5u: I/O error\n", group);
 132                 return 0;
 133         }
 134 +
 135 +       gdp = ext4_get_group_desc(sb, group, NULL);
 136 +       if (gdp != NULL)
 137 +               free = ext4_free_blks_count(sb, gdp);
 138 +
 139         ext4_lock_group(sb, group);
 140         memcpy(&sg, ext4_get_group_info(sb, group), i);
 141         ext4_unlock_group(sb, group);
 142         ext4_mb_unload_buddy(&e4b);
 143
 144 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 145 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 146 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [",
 147 +                       (long unsigned int)group, sg.info.bb_free, free,
 148 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 149 +                       sg.info.bb_prealloc_nr);
 150         for (i = 0; i <= 13; i++)
 151                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
 152                                 sg.info.bb_counters[i] : 0);
 153 @@ -3411,23 +3428,68 @@ static void ext4_mb_generate_from_freeli
 154  }
 155
 156  /*
 157 + * check free blocks in bitmap match free block in group descriptor
 158 + * do this before taking preallocated blocks into account to be able
 159 + * to detect on-disk corruptions. The group lock should be hold by the
 160 + * caller.
 161 + */
 162 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 163 +                               struct ext4_group_desc *gdp, int group)
 164 +{
 165 +       unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
 166 +       unsigned short i, first, free = 0;
 167 +
 168 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 169 +
 170 +       while (i < max) {
 171 +               first = i;
 172 +               i = mb_find_next_bit(bitmap, max, i);
 173 +               if (i > max)
 174 +                       i = max;
 175 +               free += i - first;
 176 +               if (i < max)
 177 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 178 +       }
 179 +
 180 +       if (free != ext4_free_blks_count(sb, gdp)) {
 181 +               ext4_error(sb, "on-disk bitmap for group %d"
 182 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 183 +                       group, free, ext4_free_blks_count(sb, gdp));
 184 +               return -EIO;
 185 +       }
 186 +       return 0;
 187 +}
 188 +
 189 +/*
 190   * the function goes through all preallocation in this group and marks them
 191   * used in in-core bitmap. buddy must be generated from this bitmap
 192   * Need to be called with ext4 group lock held
 193   */
 194  static noinline_for_stack
 195 -void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 196 +int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 197                                         ext4_group_t group)
 198  {
 199         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 200         struct ext4_prealloc_space *pa;
 201 +       struct ext4_group_desc *gdp;
 202         struct list_head *cur;
 203         ext4_group_t groupnr;
 204         ext4_grpblk_t start;
 205         int preallocated = 0;
 206         int count = 0;
 207 +       int skip = 0;
 208 +       int err;
 209         int len;
 210
 211 +       gdp = ext4_get_group_desc (sb, group, NULL);
 212 +       if (gdp == NULL)
 213 +               return -EIO;
 214 +
 215 +       /* before applying preallocations, check bitmap consistency */
 216 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 217 +       if (err)
 218 +               return err;
 219 +
 220         /* all form of preallocation discards first load group,
 221          * so the only competing code is preallocation use.
 222          * we don't need any locking here
 223 @@ -3443,14 +3505,23 @@ void ext4_mb_generate_from_pa(struct sup
 224                                              &groupnr, &start);
 225                 len = pa->pa_len;
 226                 spin_unlock(&pa->pa_lock);
 227 -               if (unlikely(len == 0))
 228 +               if (unlikely(len == 0)) {
 229 +                       skip++;
 230                         continue;
 231 +               }
 232                 BUG_ON(groupnr != group);
 233                 mb_set_bits(bitmap, start, len);
 234                 preallocated += len;
 235                 count++;
 236         }
 237 +       if (count + skip != grp->bb_prealloc_nr) {
 238 +               ext4_error(sb, "lost preallocations: "
 239 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 240 +                          count, grp->bb_prealloc_nr, skip);
 241 +               return -EIO;
 242 +       }
 243         mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
 244 +       return 0;
 245  }
 246
 247  static void ext4_mb_pa_callback(struct rcu_head *head)
 248 @@ -3509,6 +3580,7 @@ static void ext4_mb_put_pa(struct ext4_a
 249          */
 250         ext4_lock_group(sb, grp);
 251         list_del(&pa->pa_group_list);
 252 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 253         ext4_unlock_group(sb, grp);
 254
 255         spin_lock(pa->pa_obj_lock);
 256 @@ -3600,6 +3672,7 @@ ext4_mb_new_inode_pa(struct ext4_allocat
 257
 258         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 259         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 260 +       grp->bb_prealloc_nr++;
 261         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 262
 263         spin_lock(pa->pa_obj_lock);
 264 @@ -3661,6 +3734,7 @@ ext4_mb_new_group_pa(struct ext4_allocat
 265
 266         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 267         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 268 +       grp->bb_prealloc_nr++;
 269         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 270
 271         /*
 272 @@ -3829,6 +3903,8 @@ repeat:
 273
 274                 spin_unlock(&pa->pa_lock);
 275
 276 +               BUG_ON(grp->bb_prealloc_nr == 0);
 277 +               grp->bb_prealloc_nr--;
 278                 list_del(&pa->pa_group_list);
 279                 list_add(&pa->u.pa_tmp_list, &list);
 280         }
 281 @@ -3962,7 +4038,7 @@ repeat:
 282                 if (err) {
 283                         ext4_error(sb, "Error loading buddy information for %u",
 284                                         group);
 285 -                       continue;
 286 +                       return;
 287                 }
 288
 289                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 290 @@ -3974,6 +4050,8 @@ repeat:
 291                 }
 292
 293                 ext4_lock_group(sb, group);
 294 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 295 +               e4b.bd_info->bb_prealloc_nr--;
 296                 list_del(&pa->pa_group_list);
 297                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
 298                 ext4_unlock_group(sb, group);
 299 --- a/fs/ext4/mballoc.h
 300 +++ b/fs/ext4/mballoc.h
 301 @@ -87,7 +87,7 @@ extern u8 mb_enable_debug;
 302  /*
 303   * for which requests use 2^N search using buddies
 304   */
 305 -#define MB_DEFAULT_ORDER2_REQS         2
 306 +#define MB_DEFAULT_ORDER2_REQS         8
 307
 308  /*
 309   * default group prealloc size 512 blocks