ldiskfs/kernel_patches/patches/linux-5.8/ext4-mballoc-extra-checks.patch

   1 ---
   2  fs/ext4/ext4.h    |    1
   3  fs/ext4/mballoc.c |  103 +++++++++++++++++++++++++++++++++++++++++++++++-------
   4  fs/ext4/mballoc.h |    2 -
   5  3 files changed, 93 insertions(+), 13 deletions(-)
   6
   7 --- a/fs/ext4/ext4.h
   8 +++ b/fs/ext4/ext4.h
   9 @@ -3156,6 +3156,7 @@ 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 @@ -345,7 +345,7 @@ 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 @@ -769,7 +769,7 @@ 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 @@ -813,6 +813,7 @@ void ext4_mb_generate_buddy(struct super
  38                 grp->bb_free = free;
  39                 ext4_mark_group_bitmap_corrupted(sb, group,
  40                                         EXT4_GROUP_INFO_BBITMAP_CORRUPT);
  41 +               return -EIO;
  42         }
  43         mb_set_largest_free_order(sb, grp);
  44
  45 @@ -823,6 +824,8 @@ void ext4_mb_generate_buddy(struct super
  46         sbi->s_mb_buddies_generated++;
  47         sbi->s_mb_generation_time += period;
  48         spin_unlock(&sbi->s_bal_lock);
  49 +
  50 +       return 0;
  51  }
  52
  53  static void mb_regenerate_buddy(struct ext4_buddy *e4b)
  54 @@ -943,7 +946,7 @@ static int ext4_mb_init_cache(struct pag
  55         }
  56
  57         first_block = page->index * blocks_per_page;
  58 -       for (i = 0; i < blocks_per_page; i++) {
  59 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  60                 group = (first_block + i) >> 1;
  61                 if (group >= ngroups)
  62                         break;
  63 @@ -987,7 +990,7 @@ static int ext4_mb_init_cache(struct pag
  64                         ext4_lock_group(sb, group);
  65                         /* init the buddy */
  66                         memset(data, 0xff, blocksize);
  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 @@ -1002,7 +1005,7 @@ 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 @@ -1012,7 +1015,8 @@ static int ext4_mb_init_cache(struct pag
  82                         incore = data;
  83                 }
  84         }
  85 -       SetPageUptodate(page);
  86 +       if (likely(err == 0))
  87 +               SetPageUptodate(page);
  88
  89  out:
  90         if (bh) {
  91 @@ -2396,9 +2400,11 @@ static void *ext4_mb_seq_groups_next(str
  92  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
  93  {
  94         struct super_block *sb = PDE_DATA(file_inode(seq->file));
  95 +       struct ext4_group_desc *gdp;
  96         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
  97         int i;
  98         int err, buddy_loaded = 0;
  99 +       int free = 0;
 100         struct ext4_buddy e4b;
 101         struct ext4_group_info *grinfo;
 102         unsigned char blocksize_bits = min_t(unsigned char,
 103 @@ -2411,7 +2417,7 @@ static int ext4_mb_seq_groups_show(struc
 104
 105         group--;
 106         if (group == 0)
 107 -               seq_puts(seq, "#group: free  frags first ["
 108 +               seq_puts(seq, "#group: bfree gfree frags first pa    ["
 109                               " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
 110                               " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
 111
 112 @@ -2429,13 +2435,19 @@ static int ext4_mb_seq_groups_show(struc
 113                 buddy_loaded = 1;
 114         }
 115
 116 +       gdp = ext4_get_group_desc(sb, group, NULL);
 117 +       if (gdp != NULL)
 118 +               free = ext4_free_group_clusters(sb, gdp);
 119 +
 120         memcpy(&sg, ext4_get_group_info(sb, group), i);
 121
 122         if (buddy_loaded)
 123                 ext4_mb_unload_buddy(&e4b);
 124
 125 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 126 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 127 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [",
 128 +                       (long unsigned int)group, sg.info.bb_free, free,
 129 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 130 +                       sg.info.bb_prealloc_nr);
 131         for (i = 0; i <= 13; i++)
 132                 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
 133                                 sg.info.bb_counters[i] : 0);
 134 @@ -3742,22 +3754,71 @@ static void ext4_mb_generate_from_freeli
 135  }
 136
 137  /*
 138 + * check free blocks in bitmap match free block in group descriptor
 139 + * do this before taking preallocated blocks into account to be able
 140 + * to detect on-disk corruptions. The group lock should be hold by the
 141 + * caller.
 142 + */
 143 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 144 +                               struct ext4_group_desc *gdp, int group)
 145 +{
 146 +       unsigned short max = EXT4_CLUSTERS_PER_GROUP(sb);
 147 +       unsigned short i, first, free = 0;
 148 +       unsigned short free_in_gdp = ext4_free_group_clusters(sb, gdp);
 149 +
 150 +       if (free_in_gdp == 0 && gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
 151 +               return 0;
 152 +
 153 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 154 +
 155 +       while (i < max) {
 156 +               first = i;
 157 +               i = mb_find_next_bit(bitmap, max, i);
 158 +               if (i > max)
 159 +                       i = max;
 160 +               free += i - first;
 161 +               if (i < max)
 162 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 163 +       }
 164 +
 165 +       if (free != free_in_gdp) {
 166 +               ext4_error(sb, "on-disk bitmap for group %d"
 167 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 168 +                       group, free, free_in_gdp);
 169 +               return -EIO;
 170 +       }
 171 +       return 0;
 172 +}
 173 +
 174 +/*
 175   * the function goes through all preallocation in this group and marks them
 176   * used in in-core bitmap. buddy must be generated from this bitmap
 177   * Need to be called with ext4 group lock held
 178   */
 179  static noinline_for_stack
 180 -void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 181 +int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 182                                         ext4_group_t group)
 183  {
 184         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 185         struct ext4_prealloc_space *pa;
 186 +       struct ext4_group_desc *gdp;
 187         struct list_head *cur;
 188         ext4_group_t groupnr;
 189         ext4_grpblk_t start;
 190         int preallocated = 0;
 191 +       int skip = 0, count = 0;
 192 +       int err;
 193         int len;
 194
 195 +       gdp = ext4_get_group_desc(sb, group, NULL);
 196 +       if (gdp == NULL)
 197 +               return -EIO;
 198 +
 199 +       /* before applying preallocations, check bitmap consistency */
 200 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 201 +       if (err)
 202 +               return err;
 203 +
 204         /* all form of preallocation discards first load group,
 205          * so the only competing code is preallocation use.
 206          * we don't need any locking here
 207 @@ -3773,13 +3834,23 @@ void ext4_mb_generate_from_pa(struct sup
 208                                              &groupnr, &start);
 209                 len = pa->pa_len;
 210                 spin_unlock(&pa->pa_lock);
 211 -               if (unlikely(len == 0))
 212 +               if (unlikely(len == 0)) {
 213 +                       skip++;
 214                         continue;
 215 +               }
 216                 BUG_ON(groupnr != group);
 217                 ext4_set_bits(bitmap, start, len);
 218                 preallocated += len;
 219 +               count++;
 220 +       }
 221 +       if (count + skip != grp->bb_prealloc_nr) {
 222 +               ext4_error(sb, "lost preallocations: "
 223 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 224 +                          count, grp->bb_prealloc_nr, skip);
 225 +               return -EIO;
 226         }
 227         mb_debug(sb, "preallocated %d for group %u\n", preallocated, group);
 228 +       return 0;
 229  }
 230
 231  static void ext4_mb_pa_callback(struct rcu_head *head)
 232 @@ -3843,6 +3914,7 @@ static void ext4_mb_put_pa(struct ext4_a
 233          */
 234         ext4_lock_group(sb, grp);
 235         list_del(&pa->pa_group_list);
 236 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 237         ext4_unlock_group(sb, grp);
 238
 239         spin_lock(pa->pa_obj_lock);
 240 @@ -3934,6 +4006,7 @@ ext4_mb_new_inode_pa(struct ext4_allocat
 241         pa->pa_inode = ac->ac_inode;
 242
 243         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 244 +       grp->bb_prealloc_nr++;
 245
 246         spin_lock(pa->pa_obj_lock);
 247         list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
 248 @@ -3988,6 +4061,7 @@ ext4_mb_new_group_pa(struct ext4_allocat
 249         pa->pa_inode = NULL;
 250
 251         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 252 +       grp->bb_prealloc_nr++;
 253
 254         /*
 255          * We will later add the new pa to the right bucket
 256 @@ -4155,6 +4229,8 @@ repeat:
 257
 258                 spin_unlock(&pa->pa_lock);
 259
 260 +               BUG_ON(grp->bb_prealloc_nr == 0);
 261 +               grp->bb_prealloc_nr--;
 262                 list_del(&pa->pa_group_list);
 263                 list_add(&pa->u.pa_tmp_list, &list);
 264         }
 265 @@ -4291,7 +4367,7 @@ repeat:
 266                 if (err) {
 267                         ext4_error_err(sb, -err, "Error %d loading buddy information for %u",
 268                                        err, group);
 269 -                       continue;
 270 +                       return;
 271                 }
 272
 273                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 274 @@ -4304,6 +4380,8 @@ repeat:
 275                 }
 276
 277                 ext4_lock_group(sb, group);
 278 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 279 +               e4b.bd_info->bb_prealloc_nr--;
 280                 list_del(&pa->pa_group_list);
 281                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
 282                 ext4_unlock_group(sb, group);
 283 @@ -4598,6 +4676,7 @@ ext4_mb_discard_lg_preallocations(struct
 284                 }
 285                 ext4_lock_group(sb, group);
 286                 list_del(&pa->pa_group_list);
 287 +               ext4_get_group_info(sb, group)->bb_prealloc_nr--;
 288                 ext4_mb_release_group_pa(&e4b, pa);
 289                 ext4_unlock_group(sb, group);
 290
 291 --- a/fs/ext4/mballoc.h
 292 +++ b/fs/ext4/mballoc.h
 293 @@ -66,7 +66,7 @@
 294  /*
 295   * for which requests use 2^N search using buddies
 296   */
 297 -#define MB_DEFAULT_ORDER2_REQS         2
 298 +#define MB_DEFAULT_ORDER2_REQS         8
 299
 300  /*
 301   * default group prealloc size 512 blocks