ldiskfs/kernel_patches/patches/rhel7/ext4-mballoc-extra-checks.patch

   1 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/ext4.h
   2 ===================================================================
   3 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/ext4.h
   4 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/ext4.h
   5 @@ -2391,6 +2391,7 @@ struct ext4_group_info {
   6         ext4_grpblk_t   bb_fragments;   /* nr of freespace fragments */
   7         ext4_grpblk_t   bb_largest_free_order;/* order of largest frag in BG */
   8         struct          list_head bb_prealloc_list;
   9 +       unsigned long   bb_prealloc_nr;
  10  #ifdef DOUBLE_CHECK
  11         void            *bb_bitmap;
  12  #endif
  13 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/mballoc.c
  14 ===================================================================
  15 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/mballoc.c
  16 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/mballoc.c
  17 @@ -362,7 +362,7 @@ static const char *ext4_groupinfo_slab_n
  18         "ext4_groupinfo_64k", "ext4_groupinfo_128k"
  19  };
  20
  21 -static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  22 +static int ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
  23                                         ext4_group_t group);
  24  static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
  25                                                 ext4_group_t group);
  26 @@ -718,7 +718,7 @@ mb_set_largest_free_order(struct super_b
  27  }
  28
  29  static noinline_for_stack
  30 -void ext4_mb_generate_buddy(struct super_block *sb,
  31 +int ext4_mb_generate_buddy(struct super_block *sb,
  32                                 void *buddy, void *bitmap, ext4_group_t group)
  33  {
  34         struct ext4_group_info *grp = ext4_get_group_info(sb, group);
  35 @@ -750,19 +750,13 @@ void ext4_mb_generate_buddy(struct super
  36         grp->bb_fragments = fragments;
  37
  38         if (free != grp->bb_free) {
  39 -               ext4_grp_locked_error(sb, group, 0, 0,
  40 -                                     "block bitmap and bg descriptor "
  41 -                                     "inconsistent: %u vs %u free clusters",
  42 -                                     free, grp->bb_free);
  43 -               /*
  44 -                * If we intend to continue, we consider group descriptor
  45 -                * corrupt and update bb_free using bitmap value
  46 -                */
  47 -               grp->bb_free = free;
  48 -               if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
  49 -                       percpu_counter_sub(&sbi->s_freeclusters_counter,
  50 -                                          grp->bb_free);
  51 -               set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
  52 +               struct ext4_group_desc *gdp;
  53 +               gdp = ext4_get_group_desc(sb, group, NULL);
  54 +               ext4_error(sb, "group %lu: %u blocks in bitmap, %u in bb, "
  55 +                       "%u in gd, %lu pa's\n", (long unsigned int)group,
  56 +                       free, grp->bb_free, ext4_free_group_clusters(sb, gdp),
  57 +                       grp->bb_prealloc_nr);
  58 +               return -EIO;
  59         }
  60         mb_set_largest_free_order(sb, grp);
  61
  62 @@ -768,6 +767,8 @@ void ext4_mb_generate_buddy(struct super
  63         EXT4_SB(sb)->s_mb_buddies_generated++;
  64         EXT4_SB(sb)->s_mb_generation_time += period;
  65         spin_unlock(&EXT4_SB(sb)->s_bal_lock);
  66 +
  67 +       return 0;
  68  }
  69
  70  static void mb_regenerate_buddy(struct ext4_buddy *e4b)
  71 @@ -883,7 +884,7 @@ static int ext4_mb_init_cache(struct pag
  72         }
  73
  74         first_block = page->index * blocks_per_page;
  75 -       for (i = 0; i < blocks_per_page; i++) {
  76 +       for (i = 0; i < blocks_per_page && err == 0; i++) {
  77                 group = (first_block + i) >> 1;
  78                 if (group >= ngroups)
  79                         break;
  80 @@ -922,7 +923,7 @@ static int ext4_mb_init_cache(struct pag
  81                         ext4_lock_group(sb, group);
  82                         /* init the buddy */
  83                         memset(data, 0xff, blocksize);
  84 -                       ext4_mb_generate_buddy(sb, data, incore, group);
  85 +                       err = ext4_mb_generate_buddy(sb, data, incore, group);
  86                         ext4_unlock_group(sb, group);
  87                         incore = NULL;
  88                 } else {
  89 @@ -937,7 +938,7 @@ static int ext4_mb_init_cache(struct pag
  90                         memcpy(data, bitmap, blocksize);
  91
  92                         /* mark all preallocated blks used in in-core bitmap */
  93 -                       ext4_mb_generate_from_pa(sb, data, group);
  94 +                       err = ext4_mb_generate_from_pa(sb, data, group);
  95                         ext4_mb_generate_from_freelist(sb, data, group);
  96                         ext4_unlock_group(sb, group);
  97
  98 @@ -947,7 +948,8 @@ static int ext4_mb_init_cache(struct pag
  99                         incore = data;
 100                 }
 101         }
 102 -       SetPageUptodate(page);
 103 +       if (likely(err == 0))
 104 +               SetPageUptodate(page);
 105
 106  out:
 107         if (bh) {
 108 @@ -2224,9 +2226,11 @@ static void *ext4_mb_seq_groups_next(str
 109  static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
 110  {
 111         struct super_block *sb = seq->private;
 112 +       struct ext4_group_desc *gdp;
 113         ext4_group_t group = (ext4_group_t) ((unsigned long) v);
 114         int i;
 115         int err, buddy_loaded = 0;
 116 +       int free = 0;
 117         struct ext4_buddy e4b;
 118         struct ext4_group_info *grinfo;
 119         struct sg {
 120 @@ -2236,10 +2240,10 @@ static int ext4_mb_seq_groups_show(struc
 121
 122         group--;
 123         if (group == 0)
 124 -               seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
 125 +               seq_printf(seq, "#%-5s: %-5s %-5s %-5s %-5s %-5s"
 126                                 "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
 127                                   "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
 128 -                          "group", "free", "frags", "first",
 129 +                          "group", "bfree", "gfree", "frags", "first", "pa",
 130                            "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
 131                            "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
 132
 133 @@ -2256,13 +2260,19 @@ static int ext4_mb_seq_groups_show(struc
 134                 buddy_loaded = 1;
 135         }
 136
 137 +       gdp = ext4_get_group_desc(sb, group, NULL);
 138 +       if (gdp != NULL)
 139 +               free = ext4_free_group_clusters(sb, gdp);
 140 +
 141         memcpy(&sg, ext4_get_group_info(sb, group), i);
 142
 143         if (buddy_loaded)
 144                 ext4_mb_unload_buddy(&e4b);
 145
 146 -       seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
 147 -                       sg.info.bb_fragments, sg.info.bb_first_free);
 148 +       seq_printf(seq, "#%-5lu: %-5u %-5u %-5u %-5u %-5lu [",
 149 +                       (long unsigned int)group, sg.info.bb_free, free,
 150 +                       sg.info.bb_fragments, sg.info.bb_first_free,
 151 +                       sg.info.bb_prealloc_nr);
 152         for (i = 0; i <= 13; i++)
 153                 seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
 154                                 sg.info.bb_counters[i] : 0);
 155 @@ -3507,22 +3517,67 @@ static void ext4_mb_generate_from_freeli
 156  }
 157
 158  /*
 159 + * check free blocks in bitmap match free block in group descriptor
 160 + * do this before taking preallocated blocks into account to be able
 161 + * to detect on-disk corruptions. The group lock should be hold by the
 162 + * caller.
 163 + */
 164 +int ext4_mb_check_ondisk_bitmap(struct super_block *sb, void *bitmap,
 165 +                               struct ext4_group_desc *gdp, int group)
 166 +{
 167 +       unsigned short max = EXT4_BLOCKS_PER_GROUP(sb);
 168 +       unsigned short i, first, free = 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 != ext4_free_group_clusters(sb, 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, ext4_free_group_clusters(sb, 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 @@ -3538,13 +3593,23 @@ void ext4_mb_generate_from_pa(struct sup
 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                 ext4_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(1, "prellocated %u for group %u\n", preallocated, group);
 245 +       return 0;
 246  }
 247
 248  static void ext4_mb_pa_callback(struct rcu_head *head)
 249 @@ -3603,6 +3668,7 @@ static void ext4_mb_put_pa(struct ext4_a
 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 @@ -3697,6 +3763,7 @@ ext4_mb_new_inode_pa(struct ext4_allocat
 258
 259         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 260         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 261 +       grp->bb_prealloc_nr++;
 262         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 263
 264         spin_lock(pa->pa_obj_lock);
 265 @@ -3758,6 +3825,7 @@ ext4_mb_new_group_pa(struct ext4_allocat
 266
 267         ext4_lock_group(sb, ac->ac_b_ex.fe_group);
 268         list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 269 +       grp->bb_prealloc_nr++;
 270         ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
 271
 272         /*
 273 @@ -3927,6 +3995,8 @@ repeat:
 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 @@ -4056,7 +4126,7 @@ repeat:
 283                 if (err) {
 284                         ext4_error(sb, "Error loading buddy information for %u",
 285                                         group);
 286 -                       continue;
 287 +                       return;
 288                 }
 289
 290                 bitmap_bh = ext4_read_block_bitmap(sb, group);
 291 @@ -4068,6 +4138,8 @@ 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 @@ -4328,6 +4400,7 @@ ext4_mb_discard_lg_preallocations(struct
 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 Index: linux-3.10.0-123.el7.x86_64/fs/ext4/mballoc.h
 309 ===================================================================
 310 --- linux-3.10.0-123.el7.x86_64.orig/fs/ext4/mballoc.h
 311 +++ linux-3.10.0-123.el7.x86_64/fs/ext4/mballoc.h
 312 @@ -82,7 +82,7 @@ extern ushort ext4_mballoc_debug;
 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