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

   1 Index: linux-2.6.18-128.1.6/fs/ext4/mballoc.c
   2 ===================================================================
   3 --- linux-2.6.18-128.1.6.orig/fs/ext4/mballoc.c
   4 +++ linux-2.6.18-128.1.6/fs/ext4/mballoc.c
   5 @@ -660,7 +660,7 @@ static void ext4_mb_mark_free_simple(str
   6         }
   7  }
   8
   9 -static void ext4_mb_generate_buddy(struct super_block *sb,
  10 +static int ext4_mb_generate_buddy(struct super_block *sb,
  11                                 void *buddy, void *bitmap, ext4_group_t group)
  12  {
  13         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  14 @@ -692,14 +692,14 @@ static void ext4_mb_generate_buddy(struc
  15         grp->bb_fragments = fragments;
  16
  17         if (free != grp->bb_free) {
  18 -               ext4_error(sb, __func__,
  19 -                       "EXT4-fs: group %lu: %u blocks in bitmap, %u in gd\n",
  20 -                       group, free, grp->bb_free);
  21 -               /*
  22 -                * If we intent to continue, we consider group descritor
  23 -                * corrupt and update bb_free using bitmap value
  24 -                */
  25 -               grp->bb_free = free;
  26 +               struct ext4_group_desc *gdp;
  27 +               gdp = ext4_get_group_desc (sb, group, NULL);
  28 +               ext4_error(sb, __FUNCTION__,
  29 +                       "group %lu: %u blocks in bitmap, %u in bb, "
  30 +                       "%u in gd, %lu pa's\n", group, free, grp->bb_free,
  31 +                       le16_to_cpu(gdp->bg_free_blocks_count),
  32 +                       grp->bb_prealloc_nr);
  33 +               return -EIO;
  34         }
  35
  36         clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
  37 @@ -709,6 +709,8 @@ static void ext4_mb_generate_buddy(struc
  38         EXT4_SB(sb)->s_mb_buddies_generated++;
  39         EXT4_SB(sb)->s_mb_generation_time += period;
  40         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
  41 +
  42 +       return 0;
  43  }
  44
  45  /* The buddy information is attached the buddy cache inode
  46 @@ -814,7 +816,7 @@ static int ext4_mb_init_cache(struct pag
  47
  48         err = 0;
  49         first_block = page->index * blocks_per_page;
  50 -       for (i = 0; i < blocks_per_page; i++) {
  51 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  52                 int group;
  53                 struct ext4_group_info *grinfo;
  54
  55 @@ -848,7 +850,7 @@ static int ext4_mb_init_cache(struct pag
  56                         /*
  57                          * incore got set to the group block bitmap below
  58                          */
  59 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  60 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  61                         incore = NULL;
  62                 } else {
  63                         /* this is block of bitmap */
  64 @@ -861,7 +863,7 @@ static int ext4_mb_init_cache(struct pag
  65                         memcpy(data, bitmap, blocksize);
  66
  67                         /* mark all preallocated blks used in in-core bitmap */
  68 -                       ext4_mb_generate_from_pa(sb, data, group);
  69 +                       err = ext4_mb_generate_from_pa(sb, data, group);
  70                         ext4_unlock_group(sb, group);
  71
  72                         /* set incore so that the buddy information can be
  73 @@ -870,6 +872,7 @@ static int ext4_mb_init_cache(struct pag
  74                         incore = data;
  75                 }
  76         }
  77 +       if (likely(err == 0))
  78         SetPageUptodate(page);
  79
  80  out:
  81 @@ -1964,7 +1967,10 @@ static int ext4_mb_seq_history_show(stru
  82                         hs->result.fe_start, hs->result.fe_len);
  83                 seq_printf(seq, "%-5u %-8u %-23s free\n",
  84                                 hs->pid, hs->ino, buf2);
  85 +       } else {
  86 +               seq_printf(seq, "unknown op %d\n", hs->op);
  87         }
  88 +
  89         return 0;
  90  }
  91
  92 @@ -2092,9 +2098,11 @@ static void *ext4_mb_seq_groups_next(str
  93  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
  94  {
  95         struct super_block *sb = seq->private;
  96 +       struct ext4_group_desc *gdp;
  97         long group = (long) v;
  98         int i;
  99         int err;
 100 +       int free = 0;
 101         struct ext4_buddy e4b;
 102         struct sg {
 103                 struct ext4_group_info info;
 104 @@ -2103,10 +2111,10 @@ static int ext4_mb_seq_groups_show(struc
 105
 106         group--;
 107         if (group == 0)
 108 -               seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
 109 +               seq_printf(seq, "#%-5s: %-5s %-5s %-5s %-5s %-5s"
 110                                 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
 111                                   "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
 112 -                          "group", "free", "frags", "first",
 113 +                          "group", "free", "frags", "first", "first", "pa",
 114                            "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
 115                            "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
 116
 117 @@ -2117,13 +2125,20 @@ static int ext4_mb_seq_groups_show(struc
 118                 seq_printf(seq, "#%-5lu: I/O error\n", group);
 119                 return 0;
 120         }
 121 +
 122 +       gdp = ext4_get_group_desc(sb, group, NULL);
 123 +       if (gdp != NULL)
 124 +               free = le16_to_cpu(gdp->bg_free_blocks_count);
 125 +
 126         ext4_lock_group(sb, group);
 127         memcpy(&sg, ext4_get_group_info(sb, group), i);
 128         ext4_unlock_group(sb, group);
 129         ext4_mb_release_desc(&e4b);
 130
 131 -       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u [", group, sg.info.bb_free,
 132 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 133 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [", group,
 134 +                       sg.info.bb_free, free,
 135 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 136 +                       sg.info.bb_prealloc_nr);
 137         for (i = 0; i <= 13; i++)
 138                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
 139                                 sg.info.bb_counters[i] : 0);
 140 @@ -2226,6 +2241,7 @@ ext4_mb_store_history(struct ext4_alloca
 141         h.tail = ac->ac_tail;
 142         h.buddy = ac->ac_buddy;
 143         h.merged = 0;
 144 +       h.cr = ac->ac_criteria;
 145         if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
 146                 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
 147                                 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
 148 @@ -3539,22 +3555,67 @@ ext4_mb_use_preallocated(struct ext4_all
 149  }
 150
 151  /*
 152 + * check free blocks in bitmap match free block in group descriptor
 153 + * do this before taking preallocated blocks into account to be able
 154 + * to detect on-disk corruptions. The group lock should be hold by the
 155 + * caller.
 156 + */
 157 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 158 +                               struct ext4_group_desc *gdp, int group)
 159 +{
 160 +       unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
 161 +       unsigned short i, first, free = 0;
 162 +
 163 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 164 +
 165 +       while (i < max) {
 166 +               first = i;
 167 +               i = mb_find_next_bit(bitmap, max, i);
 168 +               if (i > max)
 169 +                       i = max;
 170 +               free += i - first;
 171 +               if (i < max)
 172 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 173 +       }
 174 +
 175 +       if (free != le16_to_cpu(gdp->bg_free_blocks_count)) {
 176 +               ext4_error(sb, __FUNCTION__, "on-disk bitmap for group %d"
 177 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 178 +                       group, free, le16_to_cpu(gdp->bg_free_blocks_count));
 179 +               return -EIO;
 180 +       }
 181 +       return 0;
 182 +}
 183 +
 184 +/*
 185   * the function goes through all preallocation in this group and marks them
 186   * used in in-core bitmap. buddy must be generated from this bitmap
 187   * Need to be called with ext4 group lock (ext4_lock_group)
 188   */
 189 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 190 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 191                                         ext4_group_t group)
 192  {
 193         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 194         struct ext4_prealloc_space *pa;
 195 +       struct ext4_group_desc *gdp;
 196         struct list_head *cur;
 197         ext4_group_t groupnr;
 198         ext4_grpblk_t start;
 199         int preallocated = 0;
 200         int count = 0;
 201 +       int skip = 0;
 202 +       int err;
 203         int len;
 204
 205 +       gdp = ext4_get_group_desc (sb, group, NULL);
 206 +       if (gdp == NULL)
 207 +               return -EIO;
 208 +
 209 +       /* before applying preallocations, check bitmap consistency */
 210 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 211 +       if (err)
 212 +               return err;
 213 +
 214         /* all form of preallocation discards first load group,
 215          * so the only competing code is preallocation use.
 216          * we don't need any locking here
 217 @@ -3570,15 +3631,24 @@ static void ext4_mb_generate_from_pa(str
 218                                              &groupnr, &start);
 219                 len = pa->pa_len;
 220                 spin_unlock(&pa->pa_lock);
 221 -               if (unlikely(len == 0))
 222 +               if (unlikely(len == 0)) {
 223 +                       skip++;
 224                         continue;
 225 +               }
 226                 BUG_ON(groupnr != group);
 227                 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
 228                                                 bitmap, start, len);
 229                 preallocated += len;
 230                 count++;
 231         }
 232 +       if (count + skip != grp->bb_prealloc_nr) {
 233 +               ext4_error(sb, __FUNCTION__, "lost preallocations: "
 234 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 235 +                          count, grp->bb_prealloc_nr, skip);
 236 +               return -EIO;
 237 +       }
 238         mb_debug("prellocated %u for group %lu\n", preallocated, group);
 239 +       return 0;
 240  }
 241
 242  static void ext4_mb_pa_callback(struct rcu_head *head)
 243 @@ -3629,6 +3699,7 @@ static void ext4_mb_put_pa(struct ext4_a
 244          */
 245         ext4_lock_group(sb, grp);
 246         list_del(&pa->pa_group_list);
 247 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 248         ext4_unlock_group(sb, grp);
 249
 250         spin_lock(pa->pa_obj_lock);
 251 @@ -3717,6 +3788,7 @@ ext4_mb_new_inode_pa(struct ext4_allocat
 252
 253         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 254         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 255 +       grp->bb_prealloc_nr++;
 256         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 257
 258         spin_lock(pa->pa_obj_lock);
 259 @@ -3776,6 +3848,7 @@ ext4_mb_new_group_pa(struct ext4_allocat
 260
 261         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 262         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 263 +       grp->bb_prealloc_nr++;
 264         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 265
 266         /*
 267 @@ -3828,6 +3901,7 @@ ext4_mb_release_inode_pa(struct ext4_bud
 268                 ac->ac_sb = sb;
 269                 ac->ac_inode = pa->pa_inode;
 270                 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
 271 +               ac->ac_o_ex.fe_len = 1;
 272         }
 273
 274         while (bit < end) {
 275 @@ -3972,6 +4046,8 @@ repeat:
 276
 277                 spin_unlock(&pa->pa_lock);
 278
 279 +               BUG_ON(grp->bb_prealloc_nr == 0);
 280 +               grp->bb_prealloc_nr--;
 281                 list_del(&pa->pa_group_list);
 282                 list_add(&pa->u.pa_tmp_list, &list);
 283         }
 284 @@ -4107,7 +4183,7 @@ repeat:
 285                 if (err) {
 286                         ext4_error(sb, __func__, "Error in loading buddy "
 287                                         "information for %lu\n", group);
 288 -                       continue;
 289 +                       return;
 290                 }
 291
 292                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 293 @@ -4119,6 +4195,8 @@ repeat:
 294                 }
 295
 296                 ext4_lock_group(sb, group);
 297 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 298 +               e4b.bd_info->bb_prealloc_nr--;
 299                 list_del(&pa->pa_group_list);
 300                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac);
 301                 ext4_unlock_group(sb, group);
 302 Index: linux-2.6.18-128.1.6/fs/ext4/mballoc.h
 303 ===================================================================
 304 --- linux-2.6.18-128.1.6.orig/fs/ext4/mballoc.h
 305 +++ linux-2.6.18-128.1.6/fs/ext4/mballoc.h
 306 @@ -119,6 +119,7 @@ struct ext4_group_info {
 307         unsigned short  bb_free;
 308         unsigned short  bb_fragments;
 309         struct          list_head bb_prealloc_list;
 310 +       unsigned long   bb_prealloc_nr;
 311  #ifdef DOUBLE_CHECK
 312         void            *bb_bitmap;
 313  #endif
 314 @@ -228,7 +229,7 @@ struct ext4_mb_history {
 315         __u16 tail;     /* what tail broke some buddy */
 316         __u16 buddy;    /* buddy the tail ^^^ broke */
 317         __u16 flags;
 318 -       __u8 cr:3;      /* which phase the result extent was found at */
 319 +       __u8 cr:8;      /* which phase the result extent was found at */
 320         __u8 op:4;
 321         __u8 merged:1;
 322  };
 323 @@ -259,7 +260,7 @@ static void ext4_mb_store_history(struct
 324
 325  struct buffer_head *read_block_bitmap(struct super_block *, ext4_group_t);
 326
 327 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 328 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 329                                         ext4_group_t group);
 330  static void ext4_mb_poll_new_transaction(struct super_block *, handle_t *);
 331  static void ext4_mb_free_committed_blocks(struct super_block *);