ldiskfs/kernel_patches/patches/ext4-mballoc-extra-checks-sles11.patch

   1 Index: linux-2.6.27.21-0.1/fs/ext4/mballoc.c
   2 ===================================================================
   3 --- linux-2.6.27.21-0.1.orig/fs/ext4/mballoc.c
   4 +++ linux-2.6.27.21-0.1/fs/ext4/mballoc.c
   5 @@ -333,7 +333,7 @@
   6  static struct kmem_cache *ext4_pspace_cachep;
   7  static struct kmem_cache *ext4_ac_cachep;
   8  static struct kmem_cache *ext4_free_ext_cachep;
   9 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  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 @@ -672,7 +672,7 @@ static void ext4_mb_mark_free_simple(str
  15         }
  16  }
  17
  18 -static void ext4_mb_generate_buddy(struct super_block *sb,
  19 +static int ext4_mb_generate_buddy(struct super_block *sb,
  20                                 void *buddy, void *bitmap, ext4_group_t group)
  21  {
  22         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  23 @@ -704,14 +704,13 @@ static void ext4_mb_generate_buddy(struc
  24         grp->bb_fragments = fragments;
  25
  26         if (free != grp->bb_free) {
  27 -               ext4_grp_locked_error(sb, group,  __func__,
  28 -                       "EXT4-fs: group %u: %u blocks in bitmap, %u in gd\n",
  29 -                       group, free, grp->bb_free);
  30 -               /*
  31 -                * If we intent to continue, we consider group descritor
  32 -                * corrupt and update bb_free using bitmap value
  33 -                */
  34 -               grp->bb_free = free;
  35 +               struct ext4_group_desc *gdp;
  36 +               gdp = ext4_get_group_desc (sb, group, NULL);
  37 +               ext4_grp_locked_error(sb, group, __func__,
  38 +                       "group %u: %u blocks in bitmap, %u in bb, "
  39 +                       "%u in gd\n", group, free, grp->bb_free,
  40 +                       ext4_free_blks_count(sb, gdp));
  41 +               return -EIO;
  42         }
  43
  44         clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
  45 @@ -721,6 +720,8 @@ static void ext4_mb_generate_buddy(struc
  46         EXT4_SB(sb)->s_mb_buddies_generated++;
  47         EXT4_SB(sb)->s_mb_generation_time += period;
  48         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
  49 +
  50 +       return 0;
  51  }
  52
  53  /* The buddy information is attached the buddy cache inode
  54 @@ -850,7 +851,7 @@ static int ext4_mb_init_cache(struct pag
  55         first_block = page->index * blocks_per_page;
  56         /* init the page  */
  57         memset(page_address(page), 0xff, PAGE_CACHE_SIZE);
  58 -       for (i = 0; i < blocks_per_page; i++) {
  59 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  60                 int group;
  61                 struct ext4_group_info *grinfo;
  62
  63 @@ -884,7 +885,7 @@ static int ext4_mb_init_cache(struct pag
  64                          * incore got set to the group block bitmap below
  65                          */
  66                         ext4_lock_group(sb, group);
  67 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  68 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  69                         ext4_unlock_group(sb, group);
  70                         incore = NULL;
  71                 } else {
  72 @@ -898,7 +899,7 @@ static int ext4_mb_init_cache(struct pag
  73                         memcpy(data, bitmap, blocksize);
  74
  75                         /* mark all preallocated blks used in in-core bitmap */
  76 -                       ext4_mb_generate_from_pa(sb, data, group);
  77 +                       err = ext4_mb_generate_from_pa(sb, data, group);
  78                         ext4_mb_generate_from_freelist(sb, data, group);
  79                         ext4_unlock_group(sb, group);
  80
  81 @@ -908,6 +909,7 @@ static int ext4_mb_init_cache(struct pag
  82                         incore = data;
  83                 }
  84         }
  85 +       if (likely(err == 0))
  86         SetPageUptodate(page);
  87
  88  out:
  89 @@ -2217,7 +2219,10 @@ static int ext4_mb_seq_history_show(stru
  90                         hs->result.fe_start, hs->result.fe_len);
  91                 seq_printf(seq, "%-5u %-8u %-23s free\n",
  92                                 hs->pid, hs->ino, buf2);
  93 +       } else {
  94 +               seq_printf(seq, "unknown op %d\n", hs->op);
  95         }
  96 +
  97         return 0;
  98  }
  99
 100 @@ -2345,9 +2350,11 @@ static void *ext4_mb_seq_groups_next(str
 101  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 102  {
 103         struct super_block *sb = seq->private;
 104 +       struct ext4_group_desc *gdp;
 105         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 106         int i;
 107         int err;
 108 +       unsigned free = 0;
 109         struct ext4_buddy e4b;
 110         struct sg {
 111                 struct ext4_group_info info;
 112 @@ -2356,10 +2363,10 @@ static int ext4_mb_seq_groups_show(struc
 113
 114         group--;
 115         if (group == 0)
 116 -               seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
 117 +               seq_printf(seq, "#%-5s: %-5s %-5s %-5s %-5s %-5s"
 118                                 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
 119                                   "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
 120 -                          "group", "free", "frags", "first",
 121 +                          "group", "free", "frags", "first", "first", "pa",
 122                            "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
 123                            "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
 124
 125 @@ -2371,12 +2378,18 @@ static int ext4_mb_seq_groups_show(struc
 126                 seq_printf(seq, "#%-5u: I/O error\n", group);
 127                 return 0;
 128         }
 129 +
 130 +       gdp = ext4_get_group_desc(sb, group, NULL);
 131 +       if (gdp != NULL)
 132 +               free = ext4_free_blks_count(sb, gdp);
 133 +
 134         ext4_lock_group(sb, group);
 135         memcpy(&sg, ext4_get_group_info(sb, group), i);
 136         ext4_unlock_group(sb, group);
 137         ext4_mb_release_desc(&e4b);
 138
 139 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 140 +       seq_printf(seq, "#%-5u: %-5u %-5u %-5u %-5u [", group,
 141 +                       sg.info.bb_free, free,
 142                         sg.info.bb_fragments, sg.info.bb_first_free);
 143         for (i = 0; i <= 13; i++)
 144                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
 145 @@ -2474,6 +2487,7 @@ ext4_mb_store_history(struct ext4_alloca
 146         h.tail = ac->ac_tail;
 147         h.buddy = ac->ac_buddy;
 148         h.merged = 0;
 149 +       h.cr = ac->ac_criteria;
 150         if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) {
 151                 if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
 152                                 ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
 153 @@ -3695,22 +3709,67 @@ 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, __FUNCTION__, "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 (ext4_lock_group)
 193   */
 194 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 195 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 196                                         ext4_group_t group)
 197  {
 198         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 199         struct ext4_prealloc_space *pa;
 200 +       struct ext4_group_desc *gdp;
 201         struct list_head *cur;
 202         ext4_group_t groupnr;
 203         ext4_grpblk_t start;
 204         int preallocated = 0;
 205         int count = 0;
 206 +       int skip = 0;
 207 +       int err;
 208         int len;
 209
 210 +       gdp = ext4_get_group_desc (sb, group, NULL);
 211 +       if (gdp == NULL)
 212 +               return -EIO;
 213 +
 214 +       /* before applying preallocations, check bitmap consistency */
 215 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 216 +       if (err)
 217 +               return err;
 218 +
 219         /* all form of preallocation discards first load group,
 220          * so the only competing code is preallocation use.
 221          * we don't need any locking here
 222 @@ -3720,8 +3778,10 @@ static void ext4_mb_generate_from_pa(str
 223                                              &groupnr, &start);
 224                 len = pa->pa_len;
 225                 spin_unlock(&pa->pa_lock);
 226 -               if (unlikely(len == 0))
 227 +               if (unlikely(len == 0)) {
 228 +                       skip++;
 229                         continue;
 230 +               }
 231                 BUG_ON(groupnr != group);
 232                 mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group),
 233                                                 bitmap, start, len);
 234 @@ -3729,6 +3789,7 @@ static void ext4_mb_generate_from_pa(str
 235                 count++;
 236         }
 237         mb_debug("prellocated %u for group %u\n", preallocated, group);
 238 +       return 0;
 239  }
 240
 241  static void ext4_mb_pa_callback(struct rcu_head *head)
 242 @@ -3978,6 +4039,7 @@ ext4_mb_release_inode_pa(struct ext4_bud
 243                 ac->ac_sb = sb;
 244                 ac->ac_inode = pa->pa_inode;
 245                 ac->ac_op = EXT4_MB_HISTORY_DISCARD;
 246 +               ac->ac_o_ex.fe_len = 1;
 247         }
 248
 249         while (bit < end) {
 250 @@ -4260,7 +4322,7 @@ repeat:
 251                         __release(e4b->alloc_semp);
 252                         ext4_error(sb, __func__, "Error in loading buddy "
 253                                         "information for %u\n", group);
 254 -                       continue;
 255 +                       return;
 256                 }
 257
 258                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 259 Index: linux-2.6.27.21-0.1/fs/ext4/mballoc.h
 260 ===================================================================
 261 --- linux-2.6.27.21-0.1.orig/fs/ext4/mballoc.h
 262 +++ linux-2.6.27.21-0.1/fs/ext4/mballoc.h
 263 @@ -219,7 +219,7 @@ struct ext4_mb_history {
 264         __u16 tail;     /* what tail broke some buddy */
 265         __u16 buddy;    /* buddy the tail ^^^ broke */
 266         __u16 flags;
 267 -       __u8 cr:3;      /* which phase the result extent was found at */
 268 +       __u8 cr:8;      /* which phase the result extent was found at */
 269         __u8 op:4;
 270         __u8 merged:1;
 271  };