ldiskfs/kernel_patches/patches/suse15/ext4-mballoc-extra-checks.patch

   1 commit f2f28f1d09c0a00b3fc569422f881931d857fac9
   2 Author:     Alex Zhuravlev <alex.zhuravlev@sun.com>
   3 AuthorDate: Tue Oct 28 17:59:09 2008 +0000
   4 Subject: ext4: detect on-disk corruption of block bitmap
   5
   6 Detect on-disk corruption of block bitmap and better checking of
   7 preallocated blocks.
   8
   9 Bugzilla-ID: b=16680
  10 Signed-off-by: Alex Zhuravlev <alex.zhuravlev@sun.com>
  11 Reviewed-by: Kalpak Shah <kalpak.shah@sun.com>
  12 Signed-off-by: Andreas Dilger <andreas.dilger@sun.com>
  13 ---
  14  fs/ext4/ext4.h    |    1
  15  fs/ext4/mballoc.c |  123 ++++++++++++++++++++++++++++++++++++++++++------------
  16  fs/ext4/mballoc.h |    2
  17  3 files changed, 99 insertions(+), 27 deletions(-)
  18
  19 --- a/fs/ext4/ext4.h
  20 +++ b/fs/ext4/ext4.h
  21 @@ -2911,6 +2911,7 @@ struct ext4_group_info {
  22         ext4_grpblk_t   bb_fragments;   /* nr of freespace fragments */
  23         ext4_grpblk_t   bb_largest_free_order;/* order of largest frag in BG */
  24         struct          list_head bb_prealloc_list;
  25 +       unsigned long   bb_prealloc_nr;
  26  #ifdef DOUBLE_CHECK
  27         void            *bb_bitmap;
  28  #endif
  29 --- a/fs/ext4/mballoc.c
  30 +++ b/fs/ext4/mballoc.c
  31 @@ -364,7 +364,7 @@ static const char * const ext4_groupinfo
  32         "ext4_groupinfo_64k", "ext4_groupinfo_128k"
  33  };
  34
  35 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  36 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  37                                         ext4_group_t group);
  38  static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
  39                                                 ext4_group_t group);
  40 @@ -720,11 +720,10 @@ mb_set_largest_free_order(struct super_b
  41  }
  42
  43  static noinline_for_stack
  44 -void ext4_mb_generate_buddy(struct super_block *sb,
  45 +int ext4_mb_generate_buddy(struct super_block *sb,
  46                                 void *buddy, void *bitmap, ext4_group_t group)
  47  {
  48         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  49 -       struct ext4_sb_info *sbi = EXT4_SB(sb);
  50         ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
  51         ext4_grpblk_t i = 0;
  52         ext4_grpblk_t first;
  53 @@ -753,19 +752,13 @@ void ext4_mb_generate_buddy(struct super
  54         grp->bb_fragments = fragments;
  55
  56         if (free != grp->bb_free) {
  57 -               ext4_grp_locked_error(sb, group, 0, 0,
  58 -                                     "block bitmap and bg descriptor "
  59 -                                     "inconsistent: %u vs %u free clusters",
  60 -                                     free, grp->bb_free);
  61 -               /*
  62 -                * If we intend to continue, we consider group descriptor
  63 -                * corrupt and update bb_free using bitmap value
  64 -                */
  65 -               grp->bb_free = free;
  66 -               if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
  67 -                       percpu_counter_sub(&sbi->s_freeclusters_counter,
  68 -                                          grp->bb_free);
  69 -               set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
  70 +               struct ext4_group_desc *gdp;
  71 +               gdp = ext4_get_group_desc(sb, group, NULL);
  72 +               ext4_error(sb, "group %lu: %u blocks in bitmap, %u in bb, "
  73 +                       "%u in gd, %lu pa's\n", (long unsigned int)group,
  74 +                       free, grp->bb_free, ext4_free_group_clusters(sb, gdp),
  75 +                       grp->bb_prealloc_nr);
  76 +               return -EIO;
  77         }
  78         mb_set_largest_free_order(sb, grp);
  79
  80 @@ -776,6 +769,8 @@ void ext4_mb_generate_buddy(struct super
  81         EXT4_SB(sb)->s_mb_buddies_generated++;
  82         EXT4_SB(sb)->s_mb_generation_time += period;
  83         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
  84 +
  85 +       return 0;
  86  }
  87
  88  static void mb_regenerate_buddy(struct ext4_buddy *e4b)
  89 @@ -896,7 +891,7 @@ static int ext4_mb_init_cache(struct pag
  90         }
  91
  92         first_block = page->index * blocks_per_page;
  93 -       for (i = 0; i < blocks_per_page; i++) {
  94 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  95                 group = (first_block + i) >> 1;
  96                 if (group >= ngroups)
  97                         break;
  98 @@ -940,7 +935,7 @@ static int ext4_mb_init_cache(struct pag
  99                         ext4_lock_group(sb, group);
 100                         /* init the buddy */
 101                         memset(data, 0xff, blocksize);
 102 -                       ext4_mb_generate_buddy(sb, data, incore, group);
 103 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
 104                         ext4_unlock_group(sb, group);
 105                         incore = NULL;
 106                 } else {
 107 @@ -955,7 +950,7 @@ static int ext4_mb_init_cache(struct pag
 108                         memcpy(data, bitmap, blocksize);
 109
 110                         /* mark all preallocated blks used in in-core bitmap */
 111 -                       ext4_mb_generate_from_pa(sb, data, group);
 112 +                       err = ext4_mb_generate_from_pa(sb, data, group);
 113                         ext4_mb_generate_from_freelist(sb, data, group);
 114                         ext4_unlock_group(sb, group);
 115
 116 @@ -965,7 +960,8 @@ static int ext4_mb_init_cache(struct pag
 117                         incore = data;
 118                 }
 119         }
 120 -       SetPageUptodate(page);
 121 +       if (likely(err == 0))
 122 +               SetPageUptodate(page);
 123
 124  out:
 125         if (bh) {
 126 @@ -2314,9 +2310,11 @@ static void *ext4_mb_seq_groups_next(str
 127  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 128  {
 129         struct super_block *sb = seq->private;
 130 +       struct ext4_group_desc *gdp;
 131         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 132         int i;
 133         int err, buddy_loaded = 0;
 134 +       int free = 0;
 135         struct ext4_buddy e4b;
 136         struct ext4_group_info *grinfo;
 137         struct sg {
 138 @@ -2326,7 +2324,7 @@ static int ext4_mb_seq_groups_show(struc
 139
 140         group--;
 141         if (group == 0)
 142 -               seq_puts(seq, "#group: free  frags first ["
 143 +               seq_puts(seq, "#group: bfree gfree frags first pa    ["
 144                               " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
 145                               " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
 146
 147 @@ -2343,13 +2341,19 @@ static int ext4_mb_seq_groups_show(struc
 148                 buddy_loaded = 1;
 149         }
 150
 151 +       gdp = ext4_get_group_desc(sb, group, NULL);
 152 +       if (gdp != NULL)
 153 +               free = ext4_free_group_clusters(sb, gdp);
 154 +
 155         memcpy(&sg, ext4_get_group_info(sb, group), i);
 156
 157         if (buddy_loaded)
 158                 ext4_mb_unload_buddy(&e4b);
 159
 160 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 161 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 162 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [",
 163 +                       (long unsigned int)group, sg.info.bb_free, free,
 164 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 165 +                       sg.info.bb_prealloc_nr);
 166         for (i = 0; i <= 13; i++)
 167                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
 168                                 sg.info.bb_counters[i] : 0);
 169 @@ -3617,22 +3621,71 @@ static void ext4_mb_generate_from_freeli
 170  }
 171
 172  /*
 173 + * check free blocks in bitmap match free block in group descriptor
 174 + * do this before taking preallocated blocks into account to be able
 175 + * to detect on-disk corruptions. The group lock should be hold by the
 176 + * caller.
 177 + */
 178 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 179 +                               struct ext4_group_desc *gdp, int group)
 180 +{
 181 +       unsigned short max = EXT4_CLUSTERS_PER_GROUP(sb);
 182 +       unsigned short i, first, free = 0;
 183 +       unsigned short free_in_gdp = ext4_free_group_clusters(sb, gdp);
 184 +
 185 +       if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
 186 +               return 0;
 187 +
 188 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 189 +
 190 +       while (i < max) {
 191 +               first = i;
 192 +               i = mb_find_next_bit(bitmap, max, i);
 193 +               if (i > max)
 194 +                       i = max;
 195 +               free += i - first;
 196 +               if (i < max)
 197 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 198 +       }
 199 +
 200 +       if (free != free_in_gdp) {
 201 +               ext4_error(sb, "on-disk bitmap for group %d"
 202 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 203 +                       group, free, free_in_gdp);
 204 +               return -EIO;
 205 +       }
 206 +       return 0;
 207 +}
 208 +
 209 +/*
 210   * the function goes through all preallocation in this group and marks them
 211   * used in in-core bitmap. buddy must be generated from this bitmap
 212   * Need to be called with ext4 group lock held
 213   */
 214  static noinline_for_stack
 215 -void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 216 +int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 217                                         ext4_group_t group)
 218  {
 219         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 220         struct ext4_prealloc_space *pa;
 221 +       struct ext4_group_desc *gdp;
 222         struct list_head *cur;
 223         ext4_group_t groupnr;
 224         ext4_grpblk_t start;
 225         int preallocated = 0;
 226 +       int skip = 0, count = 0;
 227 +       int err;
 228         int len;
 229
 230 +       gdp = ext4_get_group_desc(sb, group, NULL);
 231 +       if (gdp == NULL)
 232 +               return -EIO;
 233 +
 234 +       /* before applying preallocations, check bitmap consistency */
 235 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 236 +       if (err)
 237 +               return err;
 238 +
 239         /* all form of preallocation discards first load group,
 240          * so the only competing code is preallocation use.
 241          * we don't need any locking here
 242 @@ -3648,13 +3701,23 @@ void ext4_mb_generate_from_pa(struct sup
 243                                              &groupnr, &start);
 244                 len = pa->pa_len;
 245                 spin_unlock(&pa->pa_lock);
 246 -               if (unlikely(len == 0))
 247 +               if (unlikely(len == 0)) {
 248 +                       skip++;
 249                         continue;
 250 +               }
 251                 BUG_ON(groupnr != group);
 252                 ext4_set_bits(bitmap, start, len);
 253                 preallocated += len;
 254 +               count++;
 255 +       }
 256 +       if (count + skip != grp->bb_prealloc_nr) {
 257 +               ext4_error(sb, "lost preallocations: "
 258 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 259 +                          count, grp->bb_prealloc_nr, skip);
 260 +               return -EIO;
 261         }
 262         mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
 263 +       return 0;
 264  }
 265
 266  static void ext4_mb_pa_callback(struct rcu_head *head)
 267 @@ -3718,6 +3781,7 @@ static void ext4_mb_put_pa(struct ext4_a
 268          */
 269         ext4_lock_group(sb, grp);
 270         list_del(&pa->pa_group_list);
 271 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 272         ext4_unlock_group(sb, grp);
 273
 274         spin_lock(pa->pa_obj_lock);
 275 @@ -3812,6 +3876,7 @@ ext4_mb_new_inode_pa(struct ext4_allocat
 276
 277         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 278         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 279 +       grp->bb_prealloc_nr++;
 280         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 281
 282         spin_lock(pa->pa_obj_lock);
 283 @@ -3873,6 +3938,7 @@ ext4_mb_new_group_pa(struct ext4_allocat
 284
 285         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 286         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 287 +       grp->bb_prealloc_nr++;
 288         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 289
 290         /*
 291 @@ -4045,6 +4111,8 @@ repeat:
 292
 293                 spin_unlock(&pa->pa_lock);
 294
 295 +               BUG_ON(grp->bb_prealloc_nr == 0);
 296 +               grp->bb_prealloc_nr--;
 297                 list_del(&pa->pa_group_list);
 298                 list_add(&pa->u.pa_tmp_list, &list);
 299         }
 300 @@ -4175,7 +4243,7 @@ repeat:
 301                 if (err) {
 302                         ext4_error(sb, "Error %d loading buddy information for %u",
 303                                    err, group);
 304 -                       continue;
 305 +                       return;
 306                 }
 307
 308                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 309 @@ -4188,6 +4256,8 @@ repeat:
 310                 }
 311
 312                 ext4_lock_group(sb, group);
 313 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 314 +               e4b.bd_info->bb_prealloc_nr--;
 315                 list_del(&pa->pa_group_list);
 316                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
 317                 ext4_unlock_group(sb, group);
 318 @@ -4450,6 +4520,7 @@ ext4_mb_discard_lg_preallocations(struct
 319                 }
 320                 ext4_lock_group(sb, group);
 321                 list_del(&pa->pa_group_list);
 322 +               ext4_get_group_info(sb, group)->bb_prealloc_nr--;
 323                 ext4_mb_release_group_pa(&e4b, pa);
 324                 ext4_unlock_group(sb, group);
 325
 326 --- a/fs/ext4/mballoc.h
 327 +++ b/fs/ext4/mballoc.h
 328 @@ -69,7 +69,7 @@ do {                                                                  \
 329  /*
 330   * for which requests use 2^N search using buddies
 331   */
 332 -#define MB_DEFAULT_ORDER2_REQS         2
 333 +#define MB_DEFAULT_ORDER2_REQS         8
 334
 335  /*
 336   * default group prealloc size 512 blocks