ldiskfs/kernel_patches/patches/ext4-mballoc-extra-checks-rhel5.patch

   1 diff -rupN linux-2.6.18-128.1.6_1//fs/ext4/mballoc.c linux-2.6.18-128.1.6_2//fs/ext4/mballoc.c
   2 --- linux-2.6.18-128.1.6_1//fs/ext4/mballoc.c
   3 +++ linux-2.6.18-128.1.6_2//fs/ext4/mballoc.c
   4 @@ -360,8 +360,8 @@ static void ext4_mb_mark_free_simple(str
   5  static struct kmem_cache *ext4_pspace_cachep;
   6  static struct kmem_cache *ext4_ac_cachep;
   7  static struct kmem_cache *ext4_free_ext_cachep;
   8 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
   9 -                                       ext4_group_t group);
  10 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  11 +                                       ext4_group_t group);
  12  static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
  13                                                  ext4_group_t group);
  14
  15 @@ -660,7 +660,7 @@ static void ext4_mb_mark_free_simple(str
  16         }
  17  }
  18
  19 -static void ext4_mb_generate_buddy(struct super_block *sb,
  20 +static int ext4_mb_generate_buddy(struct super_block *sb,
  21                                 void *buddy, void *bitmap, ext4_group_t group)
  22  {
  23         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  24 @@ -692,14 +692,14 @@ static void ext4_mb_generate_buddy(struc
  25         grp->bb_fragments = fragments;
  26
  27         if (free != grp->bb_free) {
  28 -               ext4_grp_locked_error(sb, group,  __func__,
  29 -                       "EXT4-fs: group %u: %u blocks in bitmap, %u in gd",
  30 -                       group, free, grp->bb_free);
  31 -               /*
  32 -                * If we intent to continue, we consider group descritor
  33 -                * corrupt and update bb_free using bitmap value
  34 -                */
  35 -               grp->bb_free = free;
  36 +               struct ext4_group_desc *gdp;
  37 +               gdp = ext4_get_group_desc (sb, group, NULL);
  38 +               ext4_error(sb, __FUNCTION__,
  39 +                       "group %lu: %u blocks in bitmap, %u in bb, "
  40 +                       "%u in gd, %lu pa's\n", (long unsigned int)group,
  41 +                       free, grp->bb_free, ext4_free_blks_count(sb, gdp),
  42 +                       grp->bb_prealloc_nr);
  43 +               return -EIO;
  44         }
  45
  46         clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
  47 @@ -709,6 +709,8 @@ static void ext4_mb_generate_buddy(struc
  48         EXT4_SB(sb)->s_mb_buddies_generated++;
  49         EXT4_SB(sb)->s_mb_generation_time += period;
  50         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
  51 +
  52 +       return 0;
  53  }
  54
  55  /* The buddy information is attached the buddy cache inode
  56 @@ -814,7 +816,7 @@ static int ext4_mb_init_cache(struct pag
  57         first_block = page->index * blocks_per_page;
  58         /* init the page  */
  59         memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
  60 -       for (i = 0; i < blocks_per_page; i++) {
  61 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  62                 int group;
  63                 struct ext4_group_info *grinfo;
  64
  65 @@ -848,7 +850,7 @@ static int ext4_mb_init_cache(struct pag
  66                          * incore got set to the group block bitmap below
  67                          */
  68                         ext4_lock_group(sb, group);
  69 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  70 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  71                         ext4_unlock_group(sb, group);
  72                         incore = NULL;
  73                 } else {
  74 @@ -861,7 +863,7 @@ static int ext4_mb_init_cache(struct pag
  75                         memcpy(data, bitmap, blocksize);
  76
  77                         /* mark all preallocated blks used in in-core bitmap */
  78 -                       ext4_mb_generate_from_pa(sb, data, group);
  79 +                       err = ext4_mb_generate_from_pa(sb, data, group);
  80                         ext4_mb_generate_from_freelist(sb, data, group);
  81                         ext4_unlock_group(sb, group);
  82
  83 @@ -870,6 +872,7 @@ static int ext4_mb_init_cache(struct pag
  84                         incore = data;
  85                 }
  86         }
  87 +       if (likely(err == 0))
  88         SetPageUptodate(page);
  89
  90  out:
  91 @@ -1964,7 +1967,10 @@ static int ext4_mb_seq_history_show(stru
  92                         hs->result.fe_start, hs->result.fe_len);
  93                 seq_printf(seq, "%-5u %-8u %-23s free\n",
  94                                 hs->pid, hs->ino, buf2);
  95 +       } else {
  96 +               seq_printf(seq, "unknown op %d\n", hs->op);
  97         }
  98 +
  99         return 0;
 100  }
 101
 102 @@ -2092,9 +2098,11 @@ static void *ext4_mb_seq_groups_next(str
 103  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 104  {
 105         struct super_block *sb = seq->private;
 106 +       struct ext4_group_desc *gdp;
 107         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 108         int i;
 109         int err;
 110 +       int free = 0;
 111         struct ext4_buddy e4b;
 112         struct sg {
 113                 struct ext4_group_info info;
 114 @@ -2103,10 +2111,10 @@ static int ext4_mb_seq_groups_show(struc
 115
 116         group--;
 117         if (group == 0)
 118 -               seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
 119 +               seq_printf(seq, "#%-5s: %-5s %-5s %-5s %-5s %-5s"
 120                                 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
 121                                   "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
 122 -                          "group", "free", "frags", "first",
 123 +                          "group", "free", "frags", "first", "first", "pa",
 124                            "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
 125                            "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
 126
 127 @@ -2117,13 +2125,20 @@ static int ext4_mb_seq_groups_show(struc
 128                 seq_printf(seq, "#%-5lu: I/O error\n", group);
 129                 return 0;
 130         }
 131 +
 132 +       gdp = ext4_get_group_desc(sb, group, NULL);
 133 +       if (gdp != NULL)
 134 +               free = ext4_free_blks_count(sb, gdp);
 135 +
 136         ext4_lock_group(sb, group);
 137         memcpy(&sg, ext4_get_group_info(sb, group), i);
 138         ext4_unlock_group(sb, group);
 139         ext4_mb_release_desc(&e4b);
 140
 141 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 142 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 143 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [",
 144 +                       (long unsigned int)group, sg.info.bb_free, free,
 145 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 146 +                       sg.info.bb_prealloc_nr);
 147         for (i = 0; i <= 13; i++)
 148                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
 149                                 sg.info.bb_counters[i] : 0);
 150 @@ -2226,6 +2241,7 @@ ext4_mb_store_history(struct ext4_alloca
 151         h.tail = ac->ac_tail;
 152         h.buddy = ac->ac_buddy;
 153         h.merged = 0;
 154 +       h.cr = ac->ac_criteria;
 155         if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
 156                 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
 157                                 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
 158 @@ -3539,22 +3555,67 @@ ext4_mb_use_preallocated(struct ext4_all
 159  }
 160
 161  /*
 162 + * check free blocks in bitmap match free block in group descriptor
 163 + * do this before taking preallocated blocks into account to be able
 164 + * to detect on-disk corruptions. The group lock should be hold by the
 165 + * caller.
 166 + */
 167 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 168 +                               struct ext4_group_desc *gdp, int group)
 169 +{
 170 +       unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
 171 +       unsigned short i, first, free = 0;
 172 +
 173 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 174 +
 175 +       while (i < max) {
 176 +               first = i;
 177 +               i = mb_find_next_bit(bitmap, max, i);
 178 +               if (i > max)
 179 +                       i = max;
 180 +               free += i - first;
 181 +               if (i < max)
 182 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 183 +       }
 184 +
 185 +       if (free != ext4_free_blks_count(sb, gdp)) {
 186 +               ext4_error(sb, __FUNCTION__, "on-disk bitmap for group %d"
 187 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 188 +                       group, free, ext4_free_blks_count(sb, gdp));
 189 +               return -EIO;
 190 +       }
 191 +       return 0;
 192 +}
 193 +
 194 +/*
 195   * the function goes through all preallocation in this group and marks them
 196   * used in in-core bitmap. buddy must be generated from this bitmap
 197   * Need to be called with ext4 group lock (ext4_lock_group)
 198   */
 199 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 200 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 201                                         ext4_group_t group)
 202  {
 203         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 204         struct ext4_prealloc_space *pa;
 205 +       struct ext4_group_desc *gdp;
 206         struct list_head *cur;
 207         ext4_group_t groupnr;
 208         ext4_grpblk_t start;
 209         int preallocated = 0;
 210         int count = 0;
 211 +       int skip = 0;
 212 +       int err;
 213         int len;
 214
 215 +       gdp = ext4_get_group_desc (sb, group, NULL);
 216 +       if (gdp == NULL)
 217 +               return -EIO;
 218 +
 219 +       /* before applying preallocations, check bitmap consistency */
 220 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 221 +       if (err)
 222 +               return err;
 223 +
 224         /* all form of preallocation discards first load group,
 225          * so the only competing code is preallocation use.
 226          * we don't need any locking here
 227 @@ -3570,15 +3631,24 @@ static void ext4_mb_generate_from_pa(str
 228                                              &groupnr, &start);
 229                 len = pa->pa_len;
 230                 spin_unlock(&pa->pa_lock);
 231 -               if (unlikely(len == 0))
 232 +               if (unlikely(len == 0)) {
 233 +                       skip++;
 234                         continue;
 235 +               }
 236                 BUG_ON(groupnr != group);
 237                 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
 238                                                 bitmap, start, len);
 239                 preallocated += len;
 240                 count++;
 241         }
 242 +       if (count + skip != grp->bb_prealloc_nr) {
 243 +               ext4_error(sb, __FUNCTION__, "lost preallocations: "
 244 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 245 +                          count, grp->bb_prealloc_nr, skip);
 246 +               return -EIO;
 247 +       }
 248         mb_debug("prellocated %u for group %u\n", preallocated, group);
 249 +       return 0;
 250  }
 251
 252  static void ext4_mb_pa_callback(struct rcu_head *head)
 253 @@ -3629,6 +3699,7 @@ static void ext4_mb_put_pa(struct ext4_a
 254          */
 255         ext4_lock_group(sb, grp);
 256         list_del(&pa->pa_group_list);
 257 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 258         ext4_unlock_group(sb, grp);
 259
 260         spin_lock(pa->pa_obj_lock);
 261 @@ -3717,6 +3788,7 @@ ext4_mb_new_inode_pa(struct ext4_allocat
 262
 263         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 264         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 265 +       grp->bb_prealloc_nr++;
 266         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 267
 268         spin_lock(pa->pa_obj_lock);
 269 @@ -3776,6 +3848,7 @@ ext4_mb_new_group_pa(struct ext4_allocat
 270
 271         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 272         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 273 +       grp->bb_prealloc_nr++;
 274         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 275
 276         /*
 277 @@ -3828,6 +3901,7 @@ ext4_mb_release_inode_pa(struct ext4_bud
 278                 ac->ac_sb = sb;
 279                 ac->ac_inode = pa->pa_inode;
 280                 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
 281 +               ac->ac_o_ex.fe_len = 1;
 282         }
 283
 284         while (bit < end) {
 285 @@ -3972,6 +4046,8 @@ repeat:
 286
 287                 spin_unlock(&pa->pa_lock);
 288
 289 +               BUG_ON(grp->bb_prealloc_nr == 0);
 290 +               grp->bb_prealloc_nr--;
 291                 list_del(&pa->pa_group_list);
 292                 list_add(&pa->u.pa_tmp_list, &list);
 293         }
 294 @@ -4107,7 +4183,7 @@ repeat:
 295                 if (err) {
 296                         ext4_error(sb, __func__, "Error in loading buddy "
 297                                         "information for %u", group);
 298 -                       continue;
 299 +                       return;
 300                 }
 301
 302                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 303 @@ -4119,6 +4195,8 @@ repeat:
 304                 }
 305
 306                 ext4_lock_group(sb, group);
 307 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 308 +               e4b.bd_info->bb_prealloc_nr--;
 309                 list_del(&pa->pa_group_list);
 310                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
 311                 ext4_unlock_group(sb, group);
 312 @@ -4394,6 +4472,7 @@ ext4_mb_discard_lg_preallocations(struct
 313                 }
 314                 ext4_lock_group(sb, group);
 315                 list_del(&pa->pa_group_list);
 316 +               ext4_get_group_info(sb, group)->bb_prealloc_nr--;
 317                 ext4_mb_release_group_pa(&e4b, pa, ac);
 318                 ext4_unlock_group(sb, group);
 319
 320 diff -rupN linux-2.6.18-128.1.6/fs/ext4/ext4.h
 321 --- linux-2.6.18-128.1.6.orig/fs/ext4/ext4.h
 322 +++ linux-2.6.18-128.1.6/fs/ext4/ext4.h
 323 @@ -119,6 +119,7 @@ struct ext4_group_info {
 324         unsigned short  bb_free;
 325         unsigned short  bb_fragments;
 326         struct          list_head bb_prealloc_list;
 327 +       unsigned long   bb_prealloc_nr;
 328  #ifdef DOUBLE_CHECK
 329         void            *bb_bitmap;
 330  #endif
 331 Index: linux-2.6.18-128.1.6/fs/ext4/mballoc.h
 332 ===================================================================
 333 --- linux-2.6.18-128.1.6.orig/fs/ext4/mballoc.h
 334 +++ linux-2.6.18-128.1.6/fs/ext4/mballoc.h
 335 @@ -228,7 +229,7 @@ struct ext4_mb_history {
 336         __u16 tail;     /* what tail broke some buddy */
 337         __u16 buddy;    /* buddy the tail ^^^ broke */
 338         __u16 flags;
 339 -       __u8 cr:3;      /* which phase the result extent was found at */
 340 +       __u8 cr:8;      /* which phase the result extent was found at */
 341         __u8 op:4;
 342         __u8 merged:1;
 343  };