Bug in merge-recursive in virtual commit corner case

So I managed to create a fairly complex set of branches which are all merged back against each other at various points in time. Two of them have 3 merge bases according to git-merge-base. Tonight I tried to merge them together, but received the following wonderful error from git-merge-recursive:

  fatal: unable to read source tree (4b825dc642cb6eb9a060e54bf8d69288fbee4904)

For those in the know, that's the empty tree. This particular repository does not have the empty tree anywhere in it, which is why we can't read the object: it doesn't exist, and shouldn't exist.

Running `git-mktree </dev/null` to create the empty tree worked; git-merge-recursive ran through and cleanly merged the two branches. But the empty tree still shouldn't be in my repository, so a future git-prune is just going to whack it again. Yes, I know, I could anchor it with a ref (.git/refs/empty-tree) but I shouldn't need to use git-mktree and git-update-ref just to use git-merge-recursive.

I found the above error message in tree-diff.c's diff_tree_sha1 function. I threw in debugging and found that the new tree was the root tree of one branch and the base was the root tree of some other revision.

Apparently the empty tree is being created in merge-recursive.c:

   1219     if (merged_common_ancestors == NULL) {
   1220         /* if there is no common ancestor, make an empty tree */
   1221         struct tree *tree = xcalloc(1, sizeof(struct tree));
   1222
   1223         tree->object.parsed = 1;
   1224         tree->object.type = OBJ_TREE;
   1225         hash_sha1_file(NULL, 0, tree_type, tree->object.sha1);
   1226         merged_common_ancestors = make_virtual_commit(tree, "ancestor");
   1227     }

So basically this code crashes if its ever used in a repository that hasn't had a need for the empty tree before. :-(

I've been unable to create an isolated test case. I can't publish the repository that I caused the case in either.

I'm not quite sure what the fix should be here: patch diff_tree_sha1 to know when it gets the empty tree, or create the object in merge-recursive.c?

Shawn Pearce <spearce@spearce.org> writes:

I hit the same issue when I integrated Johannes's in-core merge; I originally used hash_sha1_file() but that results in objects that are supposed to be in the virtual parent unreadable when merging the real children. The key is to use write_sha1_file() to actually create the needed objects, and trust later prune to remove them.

Replace it with write_sha1_file() and you should be fine, I think.

Hi,

On Thu, 7 Dec 2006, Shawn Pearce wrote:

I think there is something else wrong. If you have three merge bases, I do not see how you could get into that particular code path.

Any possibility for me to have a private look at your repo?

Ciao, Dscho

Johannes Schindelin <Johannes.Schindelin@gmx.de> wrote:

Arrgh.  I can redo the merge and try to debug it myself; maybe I
can at least get enough details about what's going on to create a
dummy repo which exhibits the same behavior.
 

No, unfortunately for me (I wish I could let you debug this!). Far too many legal issues are involved. There is possible jail time on my part for sharing any of the files in there... ;-)

Although Junio mentioned he recently saw something like this in git.git when he merged in your xdl_merge() code. I wonder if its the same issue.

Hi,

can you send me at least the output of 

$ git-rev-list --parents branch1 branch2

preferrably packed? There is nothing you can do legally wrong, since you are not sending any data, only metadata.

Ciao, Dscho

Shawn Pearce <spearce@spearce.org> writes:

No, that is not related at all. My initial attempt (which is not in any commit that survives til today) used hash_sha1_file() on the result from xdl_merge() where I should have used write_sha1_file(), and I saw the same "oh, the required object is not there" errors from everywhere.