ldiskfs/kernel_patches/patches/linux-5.18/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 | 103 ++++++++++++++++++++++++++++++++++++++++------
  16  fs/ext4/mballoc.h |   2 +-
  17  3 files changed, 93 insertions(+), 13 deletions(-)
  18
  19 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
  20 index d7ae1b1..f4a1557 100644
  21 --- a/fs/ext4/ext4.h
  22 +++ b/fs/ext4/ext4.h
  23 @@ -3480,6 +3480,7 @@ struct ext4_group_info {
  24         ext4_grpblk_t   bb_largest_free_order;/* order of largest frag in BG */
  25         ext4_group_t    bb_group;       /* Group number */
  26         struct          list_head bb_prealloc_list;
  27 +       unsigned long   bb_prealloc_nr;
  28  #ifdef DOUBLE_CHECK
  29         void            *bb_bitmap;
  30  #endif
  31 diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
  32 index 9952908..5dbd7e7 100644
  33 --- a/fs/ext4/mballoc.c
  34 +++ b/fs/ext4/mballoc.c
  35 @@ -399,7 +399,7 @@ static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
  36         "ext4_groupinfo_64k", "ext4_groupinfo_128k"
  37  };
  38
  39 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  40 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  41                                         ext4_group_t group);
  42  static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
  43                                                 ext4_group_t group);
  44 @@ -1105,7 +1105,7 @@ mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
  45  }
  46
  47  static noinline_for_stack
  48 -void ext4_mb_generate_buddy(struct super_block *sb,
  49 +int ext4_mb_generate_buddy(struct super_block *sb,
  50                                 void *buddy, void *bitmap, ext4_group_t group)
  51  {
  52         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  53 @@ -1149,6 +1149,7 @@ void ext4_mb_generate_buddy(struct super_block *sb,
  54                 grp->bb_free = free;
  55                 ext4_mark_group_bitmap_corrupted(sb, group,
  56                                         EXT4_GROUP_INFO_BBITMAP_CORRUPT);
  57 +               return -EIO;
  58         }
  59         mb_set_largest_free_order(sb, grp);
  60
  61 @@ -1158,6 +1159,8 @@ void ext4_mb_generate_buddy(struct super_block *sb,
  62         atomic_inc(&sbi->s_mb_buddies_generated);
  63         atomic64_add(period, &sbi->s_mb_generation_time);
  64         mb_update_avg_fragment_size(sb, grp);
  65 +
  66 +       return 0;
  67  }
  68
  69  /* The buddy information is attached the buddy cache inode
  70 @@ -1260,7 +1263,7 @@ static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
  71         }
  72
  73         first_block = page->index * blocks_per_page;
  74 -       for (i = 0; i < blocks_per_page; i++) {
  75 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  76                 group = (first_block + i) >> 1;
  77                 if (group >= ngroups)
  78                         break;
  79 @@ -1304,7 +1307,7 @@ static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
  80                         ext4_lock_group(sb, group);
  81                         /* init the buddy */
  82                         memset(data, 0xff, blocksize);
  83 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  84 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  85                         ext4_unlock_group(sb, group);
  86                         incore = NULL;
  87                 } else {
  88 @@ -1319,7 +1322,7 @@ static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
  89                         memcpy(data, bitmap, blocksize);
  90
  91                         /* mark all preallocated blks used in in-core bitmap */
  92 -                       ext4_mb_generate_from_pa(sb, data, group);
  93 +                       err = ext4_mb_generate_from_pa(sb, data, group);
  94                         ext4_mb_generate_from_freelist(sb, data, group);
  95                         ext4_unlock_group(sb, group);
  96
  97 @@ -1329,7 +1332,8 @@ static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
  98                         incore = data;
  99                 }
 100         }
 101 -       SetPageUptodate(page);
 102 +       if (likely(err == 0))
 103 +               SetPageUptodate(page);
 104
 105  out:
 106         if (bh) {
 107 @@ -2858,9 +2862,11 @@ static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
 108  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 109  {
 110         struct super_block *sb = pde_data(file_inode(seq->file));
 111 +       struct ext4_group_desc *gdp;
 112         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 113         int i;
 114         int err, buddy_loaded = 0;
 115 +       int free = 0;
 116         struct ext4_buddy e4b;
 117         struct ext4_group_info *grinfo;
 118         unsigned char blocksize_bits = min_t(unsigned char,
 119 @@ -2873,7 +2879,7 @@ static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 120
 121         group--;
 122         if (group == 0)
 123 -               seq_puts(seq, "#group: free  frags first ["
 124 +               seq_puts(seq, "#group: bfree gfree frags first pa    ["
 125                               " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
 126                               " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
 127
 128 @@ -2891,13 +2897,19 @@ static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 129                 buddy_loaded = 1;
 130         }
 131
 132 +       gdp = ext4_get_group_desc(sb, group, NULL);
 133 +       if (gdp != NULL)
 134 +               free = ext4_free_group_clusters(sb, gdp);
 135 +
 136         memcpy(&sg, ext4_get_group_info(sb, group), i);
 137
 138         if (buddy_loaded)
 139                 ext4_mb_unload_buddy(&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 <= blocksize_bits + 1 ?
 149                                 sg.info.bb_counters[i] : 0);
 150 @@ -4589,23 +4601,72 @@ static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
 151         return;
 152  }
 153
 154 +/*
 155 + * check free blocks in bitmap match free block in group descriptor
 156 + * do this before taking preallocated blocks into account to be able
 157 + * to detect on-disk corruptions. The group lock should be hold by the
 158 + * caller.
 159 + */
 160 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 161 +                               struct ext4_group_desc *gdp, int group)
 162 +{
 163 +       unsigned short max = EXT4_CLUSTERS_PER_GROUP(sb);
 164 +       unsigned short i, first, free = 0;
 165 +       unsigned short free_in_gdp = ext4_free_group_clusters(sb, gdp);
 166 +
 167 +       if (free_in_gdp == 0 && gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
 168 +               return 0;
 169 +
 170 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 171 +
 172 +       while (i < max) {
 173 +               first = i;
 174 +               i = mb_find_next_bit(bitmap, max, i);
 175 +               if (i > max)
 176 +                       i = max;
 177 +               free += i - first;
 178 +               if (i < max)
 179 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 180 +       }
 181 +
 182 +       if (free != free_in_gdp) {
 183 +               ext4_error(sb, "on-disk bitmap for group %d"
 184 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 185 +                       group, free, free_in_gdp);
 186 +               return -EIO;
 187 +       }
 188 +       return 0;
 189 +}
 190 +
 191  /*
 192   * the function goes through all preallocation in this group and marks them
 193   * used in in-core bitmap. buddy must be generated from this bitmap
 194   * Need to be called with ext4 group lock held
 195   */
 196  static noinline_for_stack
 197 -void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 198 +int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 199                                         ext4_group_t group)
 200  {
 201         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 202         struct ext4_prealloc_space *pa;
 203 +       struct ext4_group_desc *gdp;
 204         struct list_head *cur;
 205         ext4_group_t groupnr;
 206         ext4_grpblk_t start;
 207         int preallocated = 0;
 208 +       int skip = 0, count = 0;
 209 +       int err;
 210         int len;
 211
 212 +       gdp = ext4_get_group_desc(sb, group, NULL);
 213 +       if (gdp == NULL)
 214 +               return -EIO;
 215 +
 216 +       /* before applying preallocations, check bitmap consistency */
 217 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 218 +       if (err)
 219 +               return err;
 220 +
 221         /* all form of preallocation discards first load group,
 222          * so the only competing code is preallocation use.
 223          * we don't need any locking here
 224 @@ -4621,13 +4682,23 @@ void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 225                                              &groupnr, &start);
 226                 len = pa->pa_len;
 227                 spin_unlock(&pa->pa_lock);
 228 -               if (unlikely(len == 0))
 229 +               if (unlikely(len == 0)) {
 230 +                       skip++;
 231                         continue;
 232 +               }
 233                 BUG_ON(groupnr != group);
 234                 mb_set_bits(bitmap, start, len);
 235                 preallocated += len;
 236 +               count++;
 237 +       }
 238 +       if (count + skip != grp->bb_prealloc_nr) {
 239 +               ext4_error(sb, "lost preallocations: "
 240 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 241 +                          count, grp->bb_prealloc_nr, skip);
 242 +               return -EIO;
 243         }
 244         mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
 245 +       return 0;
 246  }
 247
 248  static void ext4_mb_mark_pa_deleted(struct super_block *sb,
 249 @@ -4711,6 +4782,7 @@ static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
 250          */
 251         ext4_lock_group(sb, grp);
 252         list_del(&pa->pa_group_list);
 253 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 254         ext4_unlock_group(sb, grp);
 255
 256         spin_lock(pa->pa_obj_lock);
 257 @@ -4802,6 +4874,7 @@ ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
 258         pa->pa_inode = ac->ac_inode;
 259
 260         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 261 +       grp->bb_prealloc_nr++;
 262
 263         spin_lock(pa->pa_obj_lock);
 264         list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
 265 @@ -4857,6 +4930,7 @@ ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
 266         pa->pa_inode = NULL;
 267
 268         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 269 +       grp->bb_prealloc_nr++;
 270
 271         /*
 272          * We will later add the new pa to the right bucket
 273 @@ -5021,6 +5095,8 @@ ext4_mb_discard_group_preallocations(struct super_block *sb,
 274
 275                 spin_unlock(&pa->pa_lock);
 276
 277 +               BUG_ON(grp->bb_prealloc_nr == 0);
 278 +               grp->bb_prealloc_nr--;
 279                 list_del(&pa->pa_group_list);
 280                 list_add(&pa->u.pa_tmp_list, &list);
 281         }
 282 @@ -5148,7 +5224,7 @@ repeat:
 283                 if (err) {
 284                         ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
 285                                        err, group);
 286 -                       continue;
 287 +                       return;
 288                 }
 289
 290                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 291 @@ -5161,6 +5237,8 @@ repeat:
 292                 }
 293
 294                 ext4_lock_group(sb, group);
 295 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 296 +               e4b.bd_info->bb_prealloc_nr--;
 297                 list_del(&pa->pa_group_list);
 298                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
 299                 ext4_unlock_group(sb, group);
 300 @@ -5455,6 +5533,7 @@ ext4_mb_discard_lg_preallocations(struct super_block *sb,
 301                 }
 302                 ext4_lock_group(sb, group);
 303                 list_del(&pa->pa_group_list);
 304 +               ext4_get_group_info(sb, group)->bb_prealloc_nr--;
 305                 ext4_mb_release_group_pa(&e4b, pa);
 306                 ext4_unlock_group(sb, group);
 307
 308 diff --git a/fs/ext4/mballoc.h b/fs/ext4/mballoc.h
 309 index 39da92c..b6c4a30 100644
 310 --- a/fs/ext4/mballoc.h
 311 +++ b/fs/ext4/mballoc.h
 312 @@ -66,7 +66,7 @@
 313  /*
 314   * for which requests use 2^N search using buddies
 315   */
 316 -#define MB_DEFAULT_ORDER2_REQS         2
 317 +#define MB_DEFAULT_ORDER2_REQS         8
 318
 319  /*
 320   * default group prealloc size 512 blocks
 321 --
 322 2.34.1
 323