ldiskfs/kernel_patches/patches/ubuntu18/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 Index: linux-4.15.0/fs/ext4/ext4.h
  15 ===================================================================
  16 --- linux-4.15.0.orig/fs/ext4/ext4.h
  17 +++ linux-4.15.0/fs/ext4/ext4.h
  18 @@ -2874,6 +2874,7 @@ struct ext4_group_info {
  19         ext4_grpblk_t   bb_fragments;   /* nr of freespace fragments */
  20         ext4_grpblk_t   bb_largest_free_order;/* order of largest frag in BG */
  21         struct          list_head bb_prealloc_list;
  22 +       unsigned long   bb_prealloc_nr;
  23  #ifdef DOUBLE_CHECK
  24         void            *bb_bitmap;
  25  #endif
  26 Index: linux-4.15.0/fs/ext4/mballoc.c
  27 ===================================================================
  28 --- linux-4.15.0.orig/fs/ext4/mballoc.c
  29 +++ linux-4.15.0/fs/ext4/mballoc.c
  30 @@ -363,7 +363,7 @@ static const char * const ext4_groupinfo
  31         "ext4_groupinfo_64k", "ext4_groupinfo_128k"
  32  };
  33
  34 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  35 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  36                                         ext4_group_t group);
  37  static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
  38                                                 ext4_group_t group);
  39 @@ -717,7 +717,7 @@ mb_set_largest_free_order(struct super_b
  40  }
  41
  42  static noinline_for_stack
  43 -void ext4_mb_generate_buddy(struct super_block *sb,
  44 +int ext4_mb_generate_buddy(struct super_block *sb,
  45                                 void *buddy, void *bitmap, ext4_group_t group)
  46  {
  47         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  48 @@ -749,19 +749,13 @@ void ext4_mb_generate_buddy(struct super
  49         grp->bb_fragments = fragments;
  50
  51         if (free != grp->bb_free) {
  52 -               ext4_grp_locked_error(sb, group, 0, 0,
  53 -                                     "block bitmap and bg descriptor "
  54 -                                     "inconsistent: %u vs %u free clusters",
  55 -                                     free, grp->bb_free);
  56 -               /*
  57 -                * If we intend to continue, we consider group descriptor
  58 -                * corrupt and update bb_free using bitmap value
  59 -                */
  60 -               grp->bb_free = free;
  61 -               if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
  62 -                       percpu_counter_sub(&sbi->s_freeclusters_counter,
  63 -                                          grp->bb_free);
  64 -               set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
  65 +               struct ext4_group_desc *gdp;
  66 +               gdp = ext4_get_group_desc(sb, group, NULL);
  67 +               ext4_error(sb, "group %lu: %u blocks in bitmap, %u in bb, "
  68 +                       "%u in gd, %lu pa's\n", (long unsigned int)group,
  69 +                       free, grp->bb_free, ext4_free_group_clusters(sb, gdp),
  70 +                       grp->bb_prealloc_nr);
  71 +               return -EIO;
  72         }
  73         mb_set_largest_free_order(sb, grp);
  74
  75 @@ -772,6 +766,8 @@ void ext4_mb_generate_buddy(struct super
  76         EXT4_SB(sb)->s_mb_buddies_generated++;
  77         EXT4_SB(sb)->s_mb_generation_time += period;
  78         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
  79 +
  80 +       return 0;
  81  }
  82
  83  static void mb_regenerate_buddy(struct ext4_buddy *e4b)
  84 @@ -892,7 +888,7 @@ static int ext4_mb_init_cache(struct pag
  85         }
  86
  87         first_block = page->index * blocks_per_page;
  88 -       for (i = 0; i < blocks_per_page; i++) {
  89 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  90                 group = (first_block + i) >> 1;
  91                 if (group >= ngroups)
  92                         break;
  93 @@ -936,7 +932,7 @@ static int ext4_mb_init_cache(struct pag
  94                         ext4_lock_group(sb, group);
  95                         /* init the buddy */
  96                         memset(data, 0xff, blocksize);
  97 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  98 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  99                         ext4_unlock_group(sb, group);
 100                         incore = NULL;
 101                 } else {
 102 @@ -951,7 +947,7 @@ static int ext4_mb_init_cache(struct pag
 103                         memcpy(data, bitmap, blocksize);
 104
 105                         /* mark all preallocated blks used in in-core bitmap */
 106 -                       ext4_mb_generate_from_pa(sb, data, group);
 107 +                       err = ext4_mb_generate_from_pa(sb, data, group);
 108                         ext4_mb_generate_from_freelist(sb, data, group);
 109                         ext4_unlock_group(sb, group);
 110
 111 @@ -961,7 +957,8 @@ static int ext4_mb_init_cache(struct pag
 112                         incore = data;
 113                 }
 114         }
 115 -       SetPageUptodate(page);
 116 +       if (likely(err == 0))
 117 +               SetPageUptodate(page);
 118
 119  out:
 120         if (bh) {
 121 @@ -2309,9 +2306,11 @@ static void *ext4_mb_seq_groups_next(str
 122  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 123  {
 124         struct super_block *sb = seq->private;
 125 +       struct ext4_group_desc *gdp;
 126         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 127         int i;
 128         int err, buddy_loaded = 0;
 129 +       int free = 0;
 130         struct ext4_buddy e4b;
 131         struct ext4_group_info *grinfo;
 132         unsigned char blocksize_bits = min_t(unsigned char,
 133 @@ -2324,7 +2323,7 @@ static int ext4_mb_seq_groups_show(struc
 134
 135         group--;
 136         if (group == 0)
 137 -               seq_puts(seq, "#group: free  frags first ["
 138 +               seq_puts(seq, "#group: bfree gfree frags first pa    ["
 139                               " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
 140                               " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
 141
 142 @@ -2342,13 +2341,19 @@ static int ext4_mb_seq_groups_show(struc
 143                 buddy_loaded = 1;
 144         }
 145
 146 +       gdp = ext4_get_group_desc(sb, group, NULL);
 147 +       if (gdp != NULL)
 148 +               free = ext4_free_group_clusters(sb, gdp);
 149 +
 150         memcpy(&sg, ext4_get_group_info(sb, group), i);
 151
 152         if (buddy_loaded)
 153                 ext4_mb_unload_buddy(&e4b);
 154
 155 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 156 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 157 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [",
 158 +                       (long unsigned int)group, sg.info.bb_free, free,
 159 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 160 +                       sg.info.bb_prealloc_nr);
 161         for (i = 0; i <= 13; i++)
 162                 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
 163                                 sg.info.bb_counters[i] : 0);
 164 @@ -3656,22 +3661,71 @@ static void ext4_mb_generate_from_freeli
 165  }
 166
 167  /*
 168 + * check free blocks in bitmap match free block in group descriptor
 169 + * do this before taking preallocated blocks into account to be able
 170 + * to detect on-disk corruptions. The group lock should be hold by the
 171 + * caller.
 172 + */
 173 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 174 +                               struct ext4_group_desc *gdp, int group)
 175 +{
 176 +       unsigned short max = EXT4_CLUSTERS_PER_GROUP(sb);
 177 +       unsigned short i, first, free = 0;
 178 +       unsigned short free_in_gdp = ext4_free_group_clusters(sb, gdp);
 179 +
 180 +       if (free_in_gdp == 0 && gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
 181 +               return 0;
 182 +
 183 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 184 +
 185 +       while (i < max) {
 186 +               first = i;
 187 +               i = mb_find_next_bit(bitmap, max, i);
 188 +               if (i > max)
 189 +                       i = max;
 190 +               free += i - first;
 191 +               if (i < max)
 192 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 193 +       }
 194 +
 195 +       if (free != free_in_gdp) {
 196 +               ext4_error(sb, "on-disk bitmap for group %d"
 197 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 198 +                       group, free, free_in_gdp);
 199 +               return -EIO;
 200 +       }
 201 +       return 0;
 202 +}
 203 +
 204 +/*
 205   * the function goes through all preallocation in this group and marks them
 206   * used in in-core bitmap. buddy must be generated from this bitmap
 207   * Need to be called with ext4 group lock held
 208   */
 209  static noinline_for_stack
 210 -void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 211 +int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 212                                         ext4_group_t group)
 213  {
 214         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 215         struct ext4_prealloc_space *pa;
 216 +       struct ext4_group_desc *gdp;
 217         struct list_head *cur;
 218         ext4_group_t groupnr;
 219         ext4_grpblk_t start;
 220         int preallocated = 0;
 221 +       int skip = 0, count = 0;
 222 +       int err;
 223         int len;
 224
 225 +       gdp = ext4_get_group_desc(sb, group, NULL);
 226 +       if (gdp == NULL)
 227 +               return -EIO;
 228 +
 229 +       /* before applying preallocations, check bitmap consistency */
 230 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 231 +       if (err)
 232 +               return err;
 233 +
 234         /* all form of preallocation discards first load group,
 235          * so the only competing code is preallocation use.
 236          * we don't need any locking here
 237 @@ -3687,13 +3741,23 @@ void ext4_mb_generate_from_pa(struct sup
 238                                              &groupnr, &start);
 239                 len = pa->pa_len;
 240                 spin_unlock(&pa->pa_lock);
 241 -               if (unlikely(len == 0))
 242 +               if (unlikely(len == 0)) {
 243 +                       skip++;
 244                         continue;
 245 +               }
 246                 BUG_ON(groupnr != group);
 247                 ext4_set_bits(bitmap, start, len);
 248                 preallocated += len;
 249 +               count++;
 250 +       }
 251 +       if (count + skip != grp->bb_prealloc_nr) {
 252 +               ext4_error(sb, "lost preallocations: "
 253 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 254 +                          count, grp->bb_prealloc_nr, skip);
 255 +               return -EIO;
 256         }
 257         mb_debug(1, "preallocated %u for group %u\n", preallocated, group);
 258 +       return 0;
 259  }
 260
 261  static void ext4_mb_pa_callback(struct rcu_head *head)
 262 @@ -3757,6 +3821,7 @@ static void ext4_mb_put_pa(struct ext4_a
 263          */
 264         ext4_lock_group(sb, grp);
 265         list_del(&pa->pa_group_list);
 266 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 267         ext4_unlock_group(sb, grp);
 268
 269         spin_lock(pa->pa_obj_lock);
 270 @@ -3851,6 +3916,7 @@ ext4_mb_new_inode_pa(struct ext4_allocat
 271
 272         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 273         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 274 +       grp->bb_prealloc_nr++;
 275         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 276
 277         spin_lock(pa->pa_obj_lock);
 278 @@ -3912,6 +3978,7 @@ ext4_mb_new_group_pa(struct ext4_allocat
 279
 280         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 281         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 282 +       grp->bb_prealloc_nr++;
 283         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 284
 285         /*
 286 @@ -4084,6 +4151,8 @@ repeat:
 287
 288                 spin_unlock(&pa->pa_lock);
 289
 290 +               BUG_ON(grp->bb_prealloc_nr == 0);
 291 +               grp->bb_prealloc_nr--;
 292                 list_del(&pa->pa_group_list);
 293                 list_add(&pa->u.pa_tmp_list, &list);
 294         }
 295 @@ -4214,7 +4283,7 @@ repeat:
 296                 if (err) {
 297                         ext4_error(sb, "Error %d loading buddy information for %u",
 298                                    err, group);
 299 -                       continue;
 300 +                       return;
 301                 }
 302
 303                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 304 @@ -4227,6 +4296,8 @@ repeat:
 305                 }
 306
 307                 ext4_lock_group(sb, group);
 308 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 309 +               e4b.bd_info->bb_prealloc_nr--;
 310                 list_del(&pa->pa_group_list);
 311                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
 312                 ext4_unlock_group(sb, group);
 313 @@ -4489,6 +4560,7 @@ ext4_mb_discard_lg_preallocations(struct
 314                 }
 315                 ext4_lock_group(sb, group);
 316                 list_del(&pa->pa_group_list);
 317 +               ext4_get_group_info(sb, group)->bb_prealloc_nr--;
 318                 ext4_mb_release_group_pa(&e4b, pa);
 319                 ext4_unlock_group(sb, group);
 320
 321 Index: linux-4.15.0/fs/ext4/mballoc.h
 322 ===================================================================
 323 --- linux-4.15.0.orig/fs/ext4/mballoc.h
 324 +++ linux-4.15.0/fs/ext4/mballoc.h
 325 @@ -70,7 +70,7 @@ do {                                                                  \
 326  /*
 327   * for which requests use 2^N search using buddies
 328   */
 329 -#define MB_DEFAULT_ORDER2_REQS         2
 330 +#define MB_DEFAULT_ORDER2_REQS         8
 331
 332  /*
 333   * default group prealloc size 512 blocks