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

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