Overview

Today, I will share with you how JGit optimizes the configuration loading process internally on file repository. After reading this article, you will understand:

  • What is Git repository and Git config
  • Why cache is needed for config
  • File snapshot as cache
  • Invalidation: file size changed
  • Invalidation: file key changed
  • Invalidation: file content modified

This article is done based on JGit 5.5.0 release candidate 1 (v5.5.0.201909041048-rc1).

Git Repository

A Git repository holds all objects and references used for managing source code (could by any type of file, but source code is what SCM’s are typically used for). In JGit, there are two types of repositories: file repository (FileRepository.java) and distributed filesystem repository (DfsRepository.java). Today, I am going to focus only on file repository, the one we used most of the time on our computer.

In Git terms all data is stored in GIT_DIR, typically a directory called .git. A work tree is maintained unless the repository is a bare repository. Typically the .git directory is located at the root of the work dir.

Directory Usage
objects/ objects
refs/ tags and heads
config configuration
info/ more configurations

In the following sections, we are going to study the loading of file GIT_DIR/config in JGit.

Git Configuration

Git configuration is stored as files, e.g. user-scoped configuration (~/.gitconfig) or repository-scoped configuration (GIT_DIR/config). They can be modified anytime by the user or other programs. To keep the configuration up-to-date and avoid unnecessary file I/O, JGit has to find a solution. Here is how FileRepository.java handles it:

/** {@inheritDoc} */
@Override
public FileBasedConfig getConfig() {
    try {
        SystemReader.getInstance().getUserConfig();
        if (repoConfig.isOutdated()) {
            loadRepoConfig();
        }
    } catch (IOException | ConfigInvalidException e) {
        throw new RuntimeException(e);
    }
    return repoConfig;
}

It loads user-scoped configuration, then loads the repo-scoped configuration. Repository configuration is loaded lazily: it will only be reloaded if the current one in memory is outdated. But in which conditions the config file is considered as outdated? 🤔 Let’s take a look together.

File Snapshot

In JGit, a file can be cached as a FileSnapshot. It caches when a file was last read, making it possible to detect future edits. This object tracks the last modified time of a file. Later during an invocation of isModified(File), the object will return true if the file may have been modified and should be re-read from disk.

In particular, it contains the following fields to achieve the goal described above:

  1. File key
  2. File size
  3. File metadata related to changes

When Git repo config is loaded, a FileSnapshot is created for the config file using factory method FileSnapshot#saveNoConfig(File):

// org.eclipse.jgit.storage.file.FileBasedConfig#load()
final FileSnapshot newSnapshot;
newSnapshot = FileSnapshot.saveNoConfig(getFile());

Then, before reading the actual file, JGit asks if the file snapshot is modified based on 3 criteria: is size changed? is file key changed? is last modified (metadata) changed?

// org.eclipse.jgit.internal.storage.file.FileSnapshot
public boolean isModified(File path) {
    // extract metadata from `path`
    ...
    sizeChanged = isSizeChanged(currSize);
    if (sizeChanged) {
        return true;
    }
    fileKeyChanged = isFileKeyChanged(currFileKey);
    if (fileKeyChanged) {
        return true;
    }
    lastModifiedChanged = isModified(currLastModified);
    if (lastModifiedChanged) {
        return true;
    }
    return false;
}

Now, let’s see how these three invalidation criteria are implemented.

File Size Changed

// org.eclipse.jgit.internal.storage.file.FileSnapshot#isModified(File)
try {
    BasicFileAttributes fileAttributes = FS.DETECTED.fileAttributes(path);
    currSize = fileAttributes.size();
    ...
} catch (IOException e) {
    currSize = path.length();
    ...
}

JGit uses two methods to read file size. Firstly, it tries to obtain this information from file attributes, if not possible, then fallback to java.io.File#length().

File attributes are obtained using Java built-in package File NIO (java.nio.file). File size belongs to basic file attributes (BasicFileAttributes.class). It means they are common to many file systems. If any I/O error occurs, JGit will fallback to File#length(), which uses JVM native method behind the screen.

It is worth mention how JGit initializes the attribute size in FileSnapshot. By default, Java sets class attribute to 0 for primitive type long. This will be confusing face to empty file (length=0). JGit avoids this by setting the initial value to -1 when length is unknown:

// org.eclipse.jgit.internal.storage.file.FileSnaphost
/**
 * An unknown file size.
 *
 * This value is used when a comparison needs to happen purely on the lastUpdate.
 */
public static final long UNKNOWN_SIZE = -1;

File Key Changed

// org.eclipse.jgit.internal.storage.file.FileSnapshot#isModified(File)
try {
    BasicFileAttributes fileAttributes = FS.DETECTED.fileAttributes(path);
    currFileKey = getFileKey(fileAttributes);
    ...
} catch (IOException e) {
    currFileKey = MISSING_FILEKEY;
    ...
}

JGit gets the file key from file basic attributes. File key is an object that uniquely identifies the given file. On some platforms or file systems, it is possible to use an identifier, or a combination of identifiers to uniquely identify a file. Such identifiers are important for operations such as file tree traversal in file systems that support symbolic links or file systems that allow a file to be an entry in more than one directory. On UNIX file systems, for example, the device ID and inode are commonly used for such purposes.

jshell> Path p = Paths.get("/Users/mincong/github/jgit/README.md")
p ==> /Users/mincong/github/jgit/README.md

jshell> Files.readAttributes(p, BasicFileAttributes.class).fileKey()
$3 ==> (dev=1000004,ino=8608922817)

If the file key changed, JGit considers the file snapshot is invalid (outdated).

Last Modified Changed

try {
    currLastModified = fileAttributes.lastModifiedTime().toInstant();
    ...
} catch (IOException e) {
    currLastModified = Instant.ofEpochMilli(path.lastModified());
    ...
}

JGit gets the last modified time from basic file attributes or from the Path object. This timestamp indicates the last modified time of the file. If it is different from the previous one, it means that the file snapshot is outdated.

However, if the last modified time remains the same, we still cannot be sure the file is not being modified. This is called the “racy Git” problem (discovered by Pasky). If two consecutive modifications happen to the file in the same timestamp, the file snapshot that appears clean when it may not be. Here are the steps for reproduction:

  1. modify ‘foo’
  2. record file snapshot
  3. modify ‘foo’ again, in-place, without changing its size

In this case, the file is “racily clean”. The last modified time remains the same after step 3, but the file snapshot is no longer valid. More detail is described in Git documentation page: https://git-scm.com/docs/racy-git/en.

Conclusion

In this article, we saw JGit’s optimization on configuration file loading. I explained what is Git repository and its configuration at different scopes. We saw how a file is cached as FileSnapshot. Then, the 3 conditions to invalidate the file snapshot: file size changed, file key changed, or last modified time changed. Also, the corner case “racy Git” to be careful about.

Hope you enjoy this article, see you the next time!

References