ldiskfs/kernel_patches/patches/ext4-mballoc-extra-checks-rhel6.patch

   1 Index: linux-stage/fs/ext4/ext4.h
   2 ===================================================================
   3 --- linux-stage.orig/fs/ext4/ext4.h     2011-03-14 16:18:28.300241437 +0800
   4 +++ linux-stage/fs/ext4/ext4.h  2011-03-14 16:33:17.056087375 +0800
   5 @@ -1770,6 +1770,7 @@
   6         ext4_grpblk_t   bb_free;        /* total free blocks */
   7         ext4_grpblk_t   bb_fragments;   /* nr of freespace fragments */
   8         struct          list_head bb_prealloc_list;
   9 +       unsigned long   bb_prealloc_nr;
  10  #ifdef DOUBLE_CHECK
  11         void            *bb_bitmap;
  12  #endif
  13 Index: linux-stage/fs/ext4/mballoc.c
  14 ===================================================================
  15 --- linux-stage.orig/fs/ext4/mballoc.c  2011-03-14 16:18:28.336242149 +0800
  16 +++ linux-stage/fs/ext4/mballoc.c       2011-03-14 16:33:27.072292006 +0800
  17 @@ -337,7 +337,7 @@
  18  static struct kmem_cache *ext4_pspace_cachep;
  19  static struct kmem_cache *ext4_ac_cachep;
  20  static struct kmem_cache *ext4_free_ext_cachep;
  21 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  22 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  23                                         ext4_group_t group);
  24  static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
  25                                                 ext4_group_t group);
  26 @@ -659,7 +659,7 @@
  27  }
  28
  29  static noinline_for_stack
  30 -void ext4_mb_generate_buddy(struct super_block *sb,
  31 +int ext4_mb_generate_buddy(struct super_block *sb,
  32                                 void *buddy, void *bitmap, ext4_group_t group)
  33  {
  34         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  35 @@ -691,14 +691,13 @@
  36         grp->bb_fragments = fragments;
  37
  38         if (free != grp->bb_free) {
  39 -               ext4_grp_locked_error(sb, group,  __func__,
  40 -                       "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
  41 -                       group, free, grp->bb_free);
  42 -               /*
  43 -                * If we intent to continue, we consider group descritor
  44 -                * corrupt and update bb_free using bitmap value
  45 -                */
  46 -               grp->bb_free = free;
  47 +               struct ext4_group_desc *gdp;
  48 +               gdp = ext4_get_group_desc (sb, group, NULL);
  49 +               ext4_error(sb, "group %lu: %u blocks in bitmap, %u in bb, "
  50 +                       "%u in gd, %lu pa's\n", (long unsigned int)group,
  51 +                       free, grp->bb_free, ext4_free_blks_count(sb, gdp),
  52 +                       grp->bb_prealloc_nr);
  53 +               return -EIO;
  54         }
  55
  56         clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
  57 @@ -708,6 +707,8 @@
  58         EXT4_SB(sb)->s_mb_buddies_generated++;
  59         EXT4_SB(sb)->s_mb_generation_time += period;
  60         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
  61 +
  62 +       return 0;
  63  }
  64
  65  /* The buddy information is attached the buddy cache inode
  66 @@ -839,7 +840,7 @@
  67         first_block = page->index * blocks_per_page;
  68         /* init the page  */
  69         memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
  70 -       for (i = 0; i < blocks_per_page; i++) {
  71 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  72                 int group;
  73                 struct ext4_group_info *grinfo;
  74
  75 @@ -874,7 +875,7 @@
  76                          * incore got set to the group block bitmap below
  77                          */
  78                         ext4_lock_group(sb, group);
  79 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  80 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  81                         ext4_unlock_group(sb, group);
  82                         incore = NULL;
  83                 } else {
  84 @@ -888,7 +889,7 @@
  85                         memcpy(data, bitmap, blocksize);
  86
  87                         /* mark all preallocated blks used in in-core bitmap */
  88 -                       ext4_mb_generate_from_pa(sb, data, group);
  89 +                       err = ext4_mb_generate_from_pa(sb, data, group);
  90                         ext4_mb_generate_from_freelist(sb, data, group);
  91                         ext4_unlock_group(sb, group);
  92
  93 @@ -898,7 +899,8 @@
  94                         incore = data;
  95                 }
  96         }
  97 -       SetPageUptodate(page);
  98 +       if (likely(err == 0))
  99 +               SetPageUptodate(page);
 100
 101  out:
 102         if (bh) {
 103 @@ -2142,9 +2144,11 @@
 104  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 105  {
 106         struct super_block *sb = seq->private;
 107 +       struct ext4_group_desc *gdp;
 108         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 109         int i;
 110         int err;
 111 +       int free = 0;
 112         struct ext4_buddy e4b;
 113         struct sg {
 114                 struct ext4_group_info info;
 115 @@ -2153,10 +2157,10 @@
 116
 117         group--;
 118         if (group == 0)
 119 -               seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
 120 +               seq_printf(seq, "#%-5s: %-5s %-5s %-5s %-5s %-5s"
 121                                 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
 122                                   "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
 123 -                          "group", "free", "frags", "first",
 124 +                          "group", "free", "frags", "first", "first", "pa",
 125                            "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
 126                            "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
 127
 128 @@ -2167,13 +2171,20 @@
 129                 seq_printf(seq, "#%-5u: I/O error\n", group);
 130                 return 0;
 131         }
 132 +
 133 +       gdp = ext4_get_group_desc(sb, group, NULL);
 134 +       if (gdp != NULL)
 135 +               free = ext4_free_blks_count(sb, gdp);
 136 +
 137         ext4_lock_group(sb, group);
 138         memcpy(&sg, ext4_get_group_info(sb, group), i);
 139         ext4_unlock_group(sb, group);
 140         ext4_mb_release_desc(&e4b);
 141
 142 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 143 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 144 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [",
 145 +                       (long unsigned int)group, sg.info.bb_free, free,
 146 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 147 +                       sg.info.bb_prealloc_nr);
 148         for (i = 0; i <= 13; i++)
 149                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
 150                                 sg.info.bb_counters[i] : 0);
 151 @@ -3354,23 +3365,68 @@
 152  }
 153
 154  /*
 155 + * check free blocks in bitmap match free block in group descriptor
 156 + * do this before taking preallocated blocks into account to be able
 157 + * to detect on-disk corruptions. The group lock should be hold by the
 158 + * caller.
 159 + */
 160 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 161 +                               struct ext4_group_desc *gdp, int group)
 162 +{
 163 +       unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
 164 +       unsigned short i, first, free = 0;
 165 +
 166 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 167 +
 168 +       while (i < max) {
 169 +               first = i;
 170 +               i = mb_find_next_bit(bitmap, max, i);
 171 +               if (i > max)
 172 +                       i = max;
 173 +               free += i - first;
 174 +               if (i < max)
 175 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 176 +       }
 177 +
 178 +       if (free != ext4_free_blks_count(sb, gdp)) {
 179 +               ext4_error(sb, "on-disk bitmap for group %d"
 180 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 181 +                       group, free, ext4_free_blks_count(sb, gdp));
 182 +               return -EIO;
 183 +       }
 184 +       return 0;
 185 +}
 186 +
 187 +/*
 188   * the function goes through all preallocation in this group and marks them
 189   * used in in-core bitmap. buddy must be generated from this bitmap
 190   * Need to be called with ext4 group lock held
 191   */
 192  static noinline_for_stack
 193 -void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 194 +int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 195                                         ext4_group_t group)
 196  {
 197         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 198         struct ext4_prealloc_space *pa;
 199 +       struct ext4_group_desc *gdp;
 200         struct list_head *cur;
 201         ext4_group_t groupnr;
 202         ext4_grpblk_t start;
 203         int preallocated = 0;
 204         int count = 0;
 205 +       int skip = 0;
 206 +       int err;
 207         int len;
 208
 209 +       gdp = ext4_get_group_desc (sb, group, NULL);
 210 +       if (gdp == NULL)
 211 +               return -EIO;
 212 +
 213 +       /* before applying preallocations, check bitmap consistency */
 214 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 215 +       if (err)
 216 +               return err;
 217 +
 218         /* all form of preallocation discards first load group,
 219          * so the only competing code is preallocation use.
 220          * we don't need any locking here
 221 @@ -3386,14 +3442,23 @@
 222                                              &groupnr, &start);
 223                 len = pa->pa_len;
 224                 spin_unlock(&pa->pa_lock);
 225 -               if (unlikely(len == 0))
 226 +               if (unlikely(len == 0)) {
 227 +                       skip++;
 228                         continue;
 229 +               }
 230                 BUG_ON(groupnr != group);
 231                 mb_set_bits(bitmap, start, len);
 232                 preallocated += len;
 233                 count++;
 234         }
 235 +       if (count + skip != grp->bb_prealloc_nr) {
 236 +               ext4_error(sb, "lost preallocations: "
 237 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 238 +                          count, grp->bb_prealloc_nr, skip);
 239 +               return -EIO;
 240 +       }
 241         mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
 242 +       return 0;
 243  }
 244
 245  static void ext4_mb_pa_callback(struct rcu_head *head)
 246 @@ -3452,6 +3517,7 @@
 247          */
 248         ext4_lock_group(sb, grp);
 249         list_del(&pa->pa_group_list);
 250 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 251         ext4_unlock_group(sb, grp);
 252
 253         spin_lock(pa->pa_obj_lock);
 254 @@ -3543,6 +3609,7 @@
 255
 256         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 257         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 258 +       grp->bb_prealloc_nr++;
 259         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 260
 261         spin_lock(pa->pa_obj_lock);
 262 @@ -3604,6 +3671,7 @@
 263
 264         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 265         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 266 +       grp->bb_prealloc_nr++;
 267         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 268
 269         /*
 270 @@ -3802,6 +3870,8 @@
 271
 272                 spin_unlock(&pa->pa_lock);
 273
 274 +               BUG_ON(grp->bb_prealloc_nr == 0);
 275 +               grp->bb_prealloc_nr--;
 276                 list_del(&pa->pa_group_list);
 277                 list_add(&pa->u.pa_tmp_list, &list);
 278         }
 279 @@ -3942,7 +4012,7 @@
 280                 if (err) {
 281                         ext4_error(sb, "Error loading buddy information for %u",
 282                                         group);
 283 -                       continue;
 284 +                       return;
 285                 }
 286
 287                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 288 @@ -3954,6 +4024,8 @@
 289                 }
 290
 291                 ext4_lock_group(sb, group);
 292 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 293 +               e4b.bd_info->bb_prealloc_nr--;
 294                 list_del(&pa->pa_group_list);
 295                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
 296                 ext4_unlock_group(sb, group);
 297 @@ -4227,6 +4299,7 @@
 298                 }
 299                 ext4_lock_group(sb, group);
 300                 list_del(&pa->pa_group_list);
 301 +               ext4_get_group_info(sb, group)->bb_prealloc_nr--;
 302                 ext4_mb_release_group_pa(&e4b, pa, ac);
 303                 ext4_unlock_group(sb, group);
 304
 305 Index: linux-stage/fs/ext4/mballoc.h
 306 ===================================================================
 307 --- linux-stage.orig/fs/ext4/mballoc.h  2011-03-14 16:18:26.670209322 +0800
 308 +++ linux-stage/fs/ext4/mballoc.h       2011-03-14 16:32:50.859552482 +0800
 309 @@ -88,7 +88,7 @@
 310  /*
 311   * for which requests use 2^N search using buddies
 312   */
 313 -#define MB_DEFAULT_ORDER2_REQS         2
 314 +#define MB_DEFAULT_ORDER2_REQS         8
 315
 316  /*
 317   * default group prealloc size 512 blocks