Checkout Internals

Checkout has to handle a lot of different cases. It examines the differences between the target tree, the baseline tree and the working directory, plus the contents of the index, and groups files into five categories:

1. UNMODIFIED - Files that match in all places.
2. SAFE - Files where the working directory and the baseline content match that can be safely updated to the target.
3. DIRTY/MISSING - Files where the working directory differs from the baseline but there is no conflicting change with the target. One example is a file that doesn't exist in the working directory - no data would be lost as a result of writing this file. Which action will be taken with these files depends on the options you use.
4. CONFLICTS - Files where changes in the working directory conflict with changes to be applied by the target. If conflicts are found, they prevent any other modifications from being made (although there are options to override that and force the update, of course).
5. UNTRACKED/IGNORED - Files in the working directory that are untracked or ignored (i.e. only in the working directory, not the other places).

Right now, this classification is done via 3 iterators (for the three trees), with a final lookup in the index. At some point, this may move to a 4 iterator version to incorporate the index better.

The actual checkout is done in five phases (at least right now).

1. The diff between the baseline and the target tree is used as a base list of possible updates to be applied.
2. Iterate through the diff and the working directory, building a list of actions to be taken (and sending notifications about conflicts and dirty files).
3. Remove any files / directories as needed (because alphabetical iteration means that an untracked directory will end up sorted after a blob that should be checked out with the same name).
4. Update all blobs.
5. Update all submodules (after 4 in case a new .gitmodules blob was checked out)

Checkout could be driven either off a target-to-workdir diff or a baseline-to-target diff. There are pros and cons of each.

Target-to-workdir means the diff includes every file that could be modified, which simplifies bookkeeping, but the code to constantly refer back to the baseline gets complicated.

Baseline-to-target has simpler code because the diff defines the action to take, but needs special handling for untracked and ignored files, if they need to be removed.

The current checkout implementation is based on a baseline-to-target diff.

Picking Actions

The most interesting aspect of this is phase 2, picking the actions that should be taken. There are a lot of corner cases, so it may be easier to start by looking at the rules for a simple 2-iterator diff:

Key

• B1,B2,B3 - blobs with different SHAs,
• Bi - ignored blob (WD only)
• T1,T2,T3 - trees with different SHAs,
• Ti - ignored tree (WD only)
• x - nothing

Diff with 2 non-workdir iterators

Now, let's make the "New" iterator into a working directory iterator, so we replace "added" items with either untracked or ignored, like this:

Diff with non-work & workdir iterators

Note: if there is a corresponding entry in the old tree, then a working directory item won't be ignored (i.e. no Bi or Ti for tracked items).

Now, expand this to three iterators: a baseline tree, a target tree, and an actual working directory tree:

Checkout From 3 Iterators (2 not workdir, 1 workdir)

(base == old HEAD; target == what to checkout; actual == working dir)

The number is followed by ' ' if no change is needed or '+' if the case needs to write to disk or '-' if something must be deleted and '*' if there should be a delete followed by an write.

There are four tiers of safe cases:

• SAFE == completely safe to update
• SAFE+MISSING == safe except the workdir is missing the expect content
• MAYBE SAFE == safe if workdir tree matches (or is missing) baseline content, which is unknown at this point
• FORCEABLE == conflict unless FORCE is given
• DIRTY == no conflict but change is not applied unless FORCE

Some slightly unusual circumstances:

• 8 - parent dir is only deleted when file is, so parent will be left if empty even though it would be deleted if the file were present
• 11 - core git does not consider this a conflict but attempts to delete T1 and gives "unable to unlink file" error yet does not skip the rest of the operation
• 12 - without FORCE file is left deleted (i.e. not restored) so new wd is dirty (and warning message "D file" is printed), with FORCE, file is restored.
• 24 - This should be considered MAYBE SAFE since effectively it is 7 and 8 combined, but core git considers this a conflict unless forced.
• 26 - This combines two cases (1 & 25) (and also implied 8 for tree content) which are ok on their own, but core git treat this as a conflict. If not forced, this is a conflict. If forced, this actually doesn't have to write anything and leaves the new blob as an untracked file.
• 32 - This is the only case where the baseline and target values match and yet we will still write to the working directory. In all other cases, if baseline == target, we don't touch the workdir (it is either already right or is "dirty"). However, since this case also implies that a ?/B1/x case will exist as well, it can be skipped.

Cases 3, 17, 24, 26, and 29 are all considered conflicts even though none of them will require making any updates to the working directory.