ldiskfs/kernel_patches/patches/rhel7/ext4_s_max_ext_tree_depth.patch

   1 Fix ext4_ext_find_extent() to already pre-allocate ext4_ext_path[]
   2 array of the max depth instead of current depth.
   3 This will avoid racy cases of concurrent ext_depth() growth in
   4 current and unsafe implementation with ext4_ext_path[] array
   5 re-[sizing,allocation], even with more recent and related patches
   6 that will be integrated in more recent Kernels.
   7
   8 Index: linux-2.6.32-504.el6.x86_64/fs/ext4/ext4.h
   9 ===================================================================
  10 --- linux-2.6.32-504.el6.x86_64.orig/fs/ext4/ext4.h
  11 +++ linux-2.6.32-504.el6.x86_64/fs/ext4/ext4.h
  12 @@ -1147,6 +1147,9 @@
  13         unsigned long s_ext_extents;
  14  #endif
  15
  16 +       /* maximum possible extents tree depth, to be computed at mount time */
  17 +       unsigned int s_max_ext_tree_depth;
  18 +
  19         /* for buddy allocator */
  20         struct ext4_group_info ***s_group_info;
  21         struct inode *s_buddy_cache;
  22 Index: linux-2.6.32-504.el6.x86_64/fs/ext4/super.c
  23 ===================================================================
  24 --- linux-2.6.32-504.el6.x86_64.orig/fs/ext4/super.c
  25 +++ linux-2.6.32-504.el6.x86_64/fs/ext4/super.c
  26 @@ -4038,6 +4038,8 @@
  27                 if (ext4_multi_mount_protect(sb, le64_to_cpu(es->s_mmp_block)))
  28                         goto failed_mount3;
  29
  30 +       ext4_ext_init(sb); /* needed before using extent-mapped journal */
  31 +
  32         /*
  33          * The first inode we look at is the journal inode.  Don't try
  34          * root first: it may be modified in the journal!
  35 @@ -4200,7 +4202,6 @@
  36                 goto failed_mount4a;
  37         }
  38
  39 -       ext4_ext_init(sb);
  40         err = ext4_mb_init(sb);
  41         if (err) {
  42                 ext4_msg(sb, KERN_ERR, "failed to initialize mballoc (%d)",
  43 Index: linux-2.6.32-504.el6.x86_64/fs/ext4/extents.c
  44 ===================================================================
  45 --- linux-2.6.32-504.el6.x86_64.orig/fs/ext4/extents.c
  46 +++ linux-2.6.32-504.el6.x86_64/fs/ext4/extents.c
  47 @@ -699,8 +699,9 @@
  48
  49         /* account possible depth increase */
  50         if (!path) {
  51 -               path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
  52 -                               GFP_NOFS);
  53 +               path = kzalloc(sizeof(struct ext4_ext_path) *
  54 +                              (EXT4_SB(inode->i_sb)->s_max_ext_tree_depth + 1),
  55 +                              GFP_NOFS);
  56                 if (!path)
  57                         return ERR_PTR(-ENOMEM);
  58                 alloc = 1;
  59 @@ -1915,11 +1916,8 @@
  60                 /* find extent for this block */
  61                 down_read(&EXT4_I(inode)->i_data_sem);
  62
  63 -               if (path && ext_depth(inode) != depth) {
  64 -                       /* depth was changed. we have to realloc path */
  65 -                       kfree(path);
  66 -                       path = NULL;
  67 -               }
  68 +               /* path of max possible depth will be allocated during
  69 +                * first pass, so its space can be re-used for each loop */
  70
  71                 path = ext4_ext_find_extent(inode, block, path);
  72                 if (IS_ERR(path)) {
  73 @@ -2664,8 +2662,9 @@
  74                         path[k].p_block =
  75                                 le16_to_cpu(path[k].p_hdr->eh_entries)+1;
  76         } else {
  77 -               path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
  78 -                              GFP_NOFS);
  79 +               path = kzalloc(sizeof(struct ext4_ext_path) *
  80 +                              (EXT4_SB(inode->i_sb)->s_max_ext_tree_depth + 1),
  81 +                              GFP_NOFS);
  82                 if (path == NULL) {
  83                         ext4_journal_stop(handle);
  84                         return -ENOMEM;
  85 @@ -3048,13 +3034,14 @@
  86   */
  87  void ext4_ext_init(struct super_block *sb)
  88  {
  89 +       ext4_fsblk_t maxblocks;
  90 +
  91         /*
  92          * possible initialization would be here
  93          */
  94
  95         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
  96 -#if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
  97 -               printk(KERN_INFO "EXT4-fs: file extents enabled"
  98 +               printk(KERN_INFO "EXT4-fs (%s): file extents enabled"
  99  #ifdef AGGRESSIVE_TEST
 100                        ", aggressive tests"
 101  #endif
 102 @@ -3064,8 +3051,31 @@
 103  #ifdef EXTENTS_STATS
 104                        ", stats"
 105  #endif
 106 -                      "\n");
 107 -#endif
 108 +                      , sb->s_id);
 109 +               EXT4_SB(sb)->s_max_ext_tree_depth = 1;
 110 +
 111 +               maxblocks = sb->s_maxbytes / sb->s_blocksize;
 112 +
 113 +               /* 1st/root level/node of extents tree stands in i_data and
 114 +                * entries stored in tree nodes can be of type ext4_extent
 115 +                * (leaf node) or ext4_extent_idx (internal node) */
 116 +               maxblocks /= (sizeof(((struct ext4_inode_info *)0x0)->i_data) -
 117 +                             sizeof(struct ext4_extent_header)) /
 118 +                            max(sizeof(struct ext4_extent),
 119 +                                sizeof(struct ext4_extent_idx));
 120 +
 121 +               /* compute maximum extents tree depth for a fully populated
 122 +                * file of max size made of only minimal/1-block extents */
 123 +               while (maxblocks > 0) {
 124 +                       maxblocks /= (sb->s_blocksize -
 125 +                                     sizeof(struct ext4_extent_header)) /
 126 +                                    max(sizeof(struct ext4_extent),
 127 +                                        sizeof(struct ext4_extent_idx));
 128 +                       EXT4_SB(sb)->s_max_ext_tree_depth++;
 129 +               }
 130 +
 131 +               printk(", maximum tree depth=%u\n",
 132 +                      EXT4_SB(sb)->s_max_ext_tree_depth);
 133  #ifdef EXTENTS_STATS
 134                 spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
 135                 EXT4_SB(sb)->s_ext_min = 1 << 30;