ldiskfs/kernel_patches/patches/rhel8/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-4.18.0-80.1.2.el8_0/fs/ext4/ext4.h
  15 ===================================================================
  16 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/ext4.h
  17 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/ext4.h
  18 @@ -2881,6 +2881,7 @@ 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-4.18.0-80.1.2.el8_0/fs/ext4/mballoc.c
  27 ===================================================================
  28 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/mballoc.c
  29 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/mballoc.c
  30 @@ -352,7 +352,7 @@ 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 @@ -708,7 +708,7 @@ 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 @@ -752,6 +752,7 @@ void ext4_mb_generate_buddy(struct super
  49                 grp->bb_free = free;
  50                 ext4_mark_group_bitmap_corrupted(sb, group,
  51                                         EXT4_GROUP_INFO_BBITMAP_CORRUPT);
  52 +               return -EIO;
  53         }
  54         mb_set_largest_free_order(sb, grp);
  55
  56 @@ -762,6 +763,8 @@ void ext4_mb_generate_buddy(struct super
  57         sbi->s_mb_buddies_generated++;
  58         sbi->s_mb_generation_time += period;
  59         spin_unlock(&sbi->s_bal_lock);
  60 +
  61 +       return 0;
  62  }
  63
  64  static void mb_regenerate_buddy(struct ext4_buddy *e4b)
  65 @@ -882,7 +885,7 @@ static int ext4_mb_init_cache(struct pag
  66         }
  67
  68         first_block = page->index * blocks_per_page;
  69 -       for (i = 0; i < blocks_per_page; i++) {
  70 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  71                 group = (first_block + i) >> 1;
  72                 if (group >= ngroups)
  73                         break;
  74 @@ -926,7 +929,7 @@ static int ext4_mb_init_cache(struct pag
  75                         ext4_lock_group(sb, group);
  76                         /* init the buddy */
  77                         memset(data, 0xff, blocksize);
  78 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  79 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  80                         ext4_unlock_group(sb, group);
  81                         incore = NULL;
  82                 } else {
  83 @@ -941,7 +944,7 @@ static int ext4_mb_init_cache(struct pag
  84                         memcpy(data, bitmap, blocksize);
  85
  86                         /* mark all preallocated blks used in in-core bitmap */
  87 -                       ext4_mb_generate_from_pa(sb, data, group);
  88 +                       err = ext4_mb_generate_from_pa(sb, data, group);
  89                         ext4_mb_generate_from_freelist(sb, data, group);
  90                         ext4_unlock_group(sb, group);
  91
  92 @@ -951,7 +954,8 @@ static int ext4_mb_init_cache(struct pag
  93                         incore = data;
  94                 }
  95         }
  96 -       SetPageUptodate(page);
  97 +       if (likely(err == 0))
  98 +               SetPageUptodate(page);
  99
 100  out:
 101         if (bh) {
 102 @@ -2280,9 +2284,11 @@ static void *ext4_mb_seq_groups_next(str
 103  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 104  {
 105         struct super_block *sb = PDE_DATA(file_inode(seq->file));
 106 +       struct ext4_group_desc *gdp;
 107         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 108         int i;
 109         int err, buddy_loaded = 0;
 110 +       int free = 0;
 111         struct ext4_buddy e4b;
 112         struct ext4_group_info *grinfo;
 113         unsigned char blocksize_bits = min_t(unsigned char,
 114 @@ -2295,7 +2301,7 @@ static int ext4_mb_seq_groups_show(struc
 115
 116         group--;
 117         if (group == 0)
 118 -               seq_puts(seq, "#group: free  frags first ["
 119 +               seq_puts(seq, "#group: bfree gfree frags first pa    ["
 120                               " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
 121                               " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
 122
 123 @@ -2313,13 +2319,19 @@ static int ext4_mb_seq_groups_show(struc
 124                 buddy_loaded = 1;
 125         }
 126
 127 +       gdp = ext4_get_group_desc(sb, group, NULL);
 128 +       if (gdp != NULL)
 129 +               free = ext4_free_group_clusters(sb, gdp);
 130 +
 131         memcpy(&sg, ext4_get_group_info(sb, group), i);
 132
 133         if (buddy_loaded)
 134                 ext4_mb_unload_buddy(&e4b);
 135
 136 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 137 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 138 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [",
 139 +                       (long unsigned int)group, sg.info.bb_free, free,
 140 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 141 +                       sg.info.bb_prealloc_nr);
 142         for (i = 0; i <= 13; i++)
 143                 seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ?
 144                                 sg.info.bb_counters[i] : 0);
 145 @@ -3614,22 +3626,71 @@ static void ext4_mb_generate_from_freeli
 146  }
 147
 148  /*
 149 + * check free blocks in bitmap match free block in group descriptor
 150 + * do this before taking preallocated blocks into account to be able
 151 + * to detect on-disk corruptions. The group lock should be hold by the
 152 + * caller.
 153 + */
 154 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 155 +                               struct ext4_group_desc *gdp, int group)
 156 +{
 157 +       unsigned short max = EXT4_CLUSTERS_PER_GROUP(sb);
 158 +       unsigned short i, first, free = 0;
 159 +       unsigned short free_in_gdp = ext4_free_group_clusters(sb, gdp);
 160 +
 161 +       if (free_in_gdp == 0 && gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))
 162 +               return 0;
 163 +
 164 +       i = mb_find_next_zero_bit(bitmap, max, 0);
 165 +
 166 +       while (i < max) {
 167 +               first = i;
 168 +               i = mb_find_next_bit(bitmap, max, i);
 169 +               if (i > max)
 170 +                       i = max;
 171 +               free += i - first;
 172 +               if (i < max)
 173 +                       i = mb_find_next_zero_bit(bitmap, max, i);
 174 +       }
 175 +
 176 +       if (free != free_in_gdp) {
 177 +               ext4_error(sb, "on-disk bitmap for group %d"
 178 +                       "corrupted: %u blocks free in bitmap, %u - in gd\n",
 179 +                       group, free, free_in_gdp);
 180 +               return -EIO;
 181 +       }
 182 +       return 0;
 183 +}
 184 +
 185 +/*
 186   * the function goes through all preallocation in this group and marks them
 187   * used in in-core bitmap. buddy must be generated from this bitmap
 188   * Need to be called with ext4 group lock held
 189   */
 190  static noinline_for_stack
 191 -void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 192 +int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 193                                         ext4_group_t group)
 194  {
 195         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 196         struct ext4_prealloc_space *pa;
 197 +       struct ext4_group_desc *gdp;
 198         struct list_head *cur;
 199         ext4_group_t groupnr;
 200         ext4_grpblk_t start;
 201         int preallocated = 0;
 202 +       int skip = 0, count = 0;
 203 +       int err;
 204         int len;
 205
 206 +       gdp = ext4_get_group_desc(sb, group, NULL);
 207 +       if (gdp == NULL)
 208 +               return -EIO;
 209 +
 210 +       /* before applying preallocations, check bitmap consistency */
 211 +       err = ext4_mb_check_ondisk_bitmap(sb, bitmap, gdp, group);
 212 +       if (err)
 213 +               return err;
 214 +
 215         /* all form of preallocation discards first load group,
 216          * so the only competing code is preallocation use.
 217          * we don't need any locking here
 218 @@ -3645,13 +3706,23 @@ void ext4_mb_generate_from_pa(struct sup
 219                                              &groupnr, &start);
 220                 len = pa->pa_len;
 221                 spin_unlock(&pa->pa_lock);
 222 -               if (unlikely(len == 0))
 223 +               if (unlikely(len == 0)) {
 224 +                       skip++;
 225                         continue;
 226 +               }
 227                 BUG_ON(groupnr != group);
 228                 ext4_set_bits(bitmap, start, len);
 229                 preallocated += len;
 230 +               count++;
 231 +       }
 232 +       if (count + skip != grp->bb_prealloc_nr) {
 233 +               ext4_error(sb, "lost preallocations: "
 234 +                          "count %d, bb_prealloc_nr %lu, skip %d\n",
 235 +                          count, grp->bb_prealloc_nr, skip);
 236 +               return -EIO;
 237         }
 238         mb_debug(1, "preallocated %u for group %u\n", preallocated, group);
 239 +       return 0;
 240  }
 241
 242  static void ext4_mb_pa_callback(struct rcu_head *head)
 243 @@ -3715,6 +3786,7 @@ static void ext4_mb_put_pa(struct ext4_a
 244          */
 245         ext4_lock_group(sb, grp);
 246         list_del(&pa->pa_group_list);
 247 +       ext4_get_group_info(sb, grp)->bb_prealloc_nr--;
 248         ext4_unlock_group(sb, grp);
 249
 250         spin_lock(pa->pa_obj_lock);
 251 @@ -3809,6 +3881,7 @@ ext4_mb_new_inode_pa(struct ext4_allocat
 252
 253         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 254         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 255 +       grp->bb_prealloc_nr++;
 256         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 257
 258         spin_lock(pa->pa_obj_lock);
 259 @@ -3870,6 +3943,7 @@ ext4_mb_new_group_pa(struct ext4_allocat
 260
 261         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 262         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 263 +       grp->bb_prealloc_nr++;
 264         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 265
 266         /*
 267 @@ -4041,6 +4115,8 @@ repeat:
 268
 269                 spin_unlock(&pa->pa_lock);
 270
 271 +               BUG_ON(grp->bb_prealloc_nr == 0);
 272 +               grp->bb_prealloc_nr--;
 273                 list_del(&pa->pa_group_list);
 274                 list_add(&pa->u.pa_tmp_list, &list);
 275         }
 276 @@ -4171,7 +4247,7 @@ repeat:
 277                 if (err) {
 278                         ext4_error(sb, "Error %d loading buddy information for %u",
 279                                    err, group);
 280 -                       continue;
 281 +                       return;
 282                 }
 283
 284                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 285 @@ -4184,6 +4260,8 @@ repeat:
 286                 }
 287
 288                 ext4_lock_group(sb, group);
 289 +               BUG_ON(e4b.bd_info->bb_prealloc_nr == 0);
 290 +               e4b.bd_info->bb_prealloc_nr--;
 291                 list_del(&pa->pa_group_list);
 292                 ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
 293                 ext4_unlock_group(sb, group);
 294 @@ -4446,6 +4524,7 @@ ext4_mb_discard_lg_preallocations(struct
 295                 }
 296                 ext4_lock_group(sb, group);
 297                 list_del(&pa->pa_group_list);
 298 +               ext4_get_group_info(sb, group)->bb_prealloc_nr--;
 299                 ext4_mb_release_group_pa(&e4b, pa);
 300                 ext4_unlock_group(sb, group);
 301
 302 Index: linux-4.18.0-80.1.2.el8_0/fs/ext4/mballoc.h
 303 ===================================================================
 304 --- linux-4.18.0-80.1.2.el8_0.orig/fs/ext4/mballoc.h
 305 +++ linux-4.18.0-80.1.2.el8_0/fs/ext4/mballoc.h
 306 @@ -70,7 +70,7 @@ do {                                                                  \
 307  /*
 308   * for which requests use 2^N search using buddies
 309   */
 310 -#define MB_DEFAULT_ORDER2_REQS         2
 311 +#define MB_DEFAULT_ORDER2_REQS         8
 312
 313  /*
 314   * default group prealloc size 512 blocks