)]}'
{
  "log": [
    {
      "commit": "2c47e605a91dde6b0514f689645e7ab336c8592a",
      "tree": "d541e61440f06f37d17e2d14cbaa78fc0cd41ff2",
      "parents": [
        "a970b0a16cc416a509d5ae8b1d70978664e6f4fe"
      ],
      "author": {
        "name": "Yan Zheng",
        "email": "zheng.yan@oracle.com",
        "time": "Sat Jun 27 21:07:35 2009 -0400"
      },
      "committer": {
        "name": "Chris Mason",
        "email": "chris.mason@oracle.com",
        "time": "Thu Jul 02 13:41:17 2009 -0400"
      },
      "message": "Btrfs: update backrefs while dropping snapshot\n\nThe new backref format has restriction on type of backref item.  If a tree\nblock isn\u0027t referenced by its owner tree, full backrefs must be used for the\npointers in it. When a tree block loses its owner tree\u0027s reference, backrefs\nfor the pointers in it should be updated to full backrefs. Current\nbtrfs_drop_snapshot misses the code that updates backrefs, so it\u0027s unsafe for\ngeneral use.\n\nThis patch adds backrefs update code to btrfs_drop_snapshot.  It isn\u0027t a\nproblem in the restricted form btrfs_drop_snapshot is used today, but for\ngeneral snapshot deletion this update is required.\n\nSigned-off-by: Yan Zheng \u003czheng.yan@oracle.com\u003e\nSigned-off-by: Chris Mason \u003cchris.mason@oracle.com\u003e\n"
    },
    {
      "commit": "5d4f98a28c7d334091c1b7744f48a1acdd2a4ae0",
      "tree": "c611d7d824cbcdb777dd2d8e33e2ed1c5df8a9c6",
      "parents": [
        "5c939df56c3ea018b58e5aa76181284c2053d699"
      ],
      "author": {
        "name": "Yan Zheng",
        "email": "zheng.yan@oracle.com",
        "time": "Wed Jun 10 10:45:14 2009 -0400"
      },
      "committer": {
        "name": "Chris Mason",
        "email": "chris.mason@oracle.com",
        "time": "Wed Jun 10 11:29:46 2009 -0400"
      },
      "message": "Btrfs: Mixed back reference  (FORWARD ROLLING FORMAT CHANGE)\n\nThis commit introduces a new kind of back reference for btrfs metadata.\nOnce a filesystem has been mounted with this commit, IT WILL NO LONGER\nBE MOUNTABLE BY OLDER KERNELS.\n\nWhen a tree block in subvolume tree is cow\u0027d, the reference counts of all\nextents it points to are increased by one.  At transaction commit time,\nthe old root of the subvolume is recorded in a \"dead root\" data structure,\nand the btree it points to is later walked, dropping reference counts\nand freeing any blocks where the reference count goes to 0.\n\nThe increments done during cow and decrements done after commit cancel out,\nand the walk is a very expensive way to go about freeing the blocks that\nare no longer referenced by the new btree root.  This commit reduces the\ntransaction overhead by avoiding the need for dead root records.\n\nWhen a non-shared tree block is cow\u0027d, we free the old block at once, and the\nnew block inherits old block\u0027s references. When a tree block with reference\ncount \u003e 1 is cow\u0027d, we increase the reference counts of all extents\nthe new block points to by one, and decrease the old block\u0027s reference count by\none.\n\nThis dead tree avoidance code removes the need to modify the reference\ncounts of lower level extents when a non-shared tree block is cow\u0027d.\nBut we still need to update back ref for all pointers in the block.\nThis is because the location of the block is recorded in the back ref\nitem.\n\nWe can solve this by introducing a new type of back ref. The new\nback ref provides information about pointer\u0027s key, level and in which\ntree the pointer lives. This information allow us to find the pointer\nby searching the tree. The shortcoming of the new back ref is that it\nonly works for pointers in tree blocks referenced by their owner trees.\n\nThis is mostly a problem for snapshots, where resolving one of these\nfuzzy back references would be O(number_of_snapshots) and quite slow.\nThe solution used here is to use the fuzzy back references in the common\ncase where a given tree block is only referenced by one root,\nand use the full back references when multiple roots have a reference\non a given block.\n\nThis commit adds per subvolume red-black tree to keep trace of cached\ninodes. The red-black tree helps the balancing code to find cached\ninodes whose inode numbers within a given range.\n\nThis commit improves the balancing code by introducing several data\nstructures to keep the state of balancing. The most important one\nis the back ref cache. It caches how the upper level tree blocks are\nreferenced. This greatly reduce the overhead of checking back ref.\n\nThe improved balancing code scales significantly better with a large\nnumber of snapshots.\n\nThis is a very large commit and was written in a number of\npieces.  But, they depend heavily on the disk format change and were\nsquashed together to make sure git bisect didn\u0027t end up in a\nbad state wrt space balancing or the format change.\n\nSigned-off-by: Yan Zheng \u003czheng.yan@oracle.com\u003e\nSigned-off-by: Chris Mason \u003cchris.mason@oracle.com\u003e\n"
    }
  ]
}