* Stop ignoring global Mercurial configuration for non-server commands
There was previously a weird duality where a simple push to or pull from a
repository hosted by SCM-Manager would load all of the system and user level
config files, but they were ignored when importing a repository. I found this
out when importing an LFS repo, and it eventually failed with a message that the
LFS extension needs to be enabled, even though it was enabled globally. The
global config should typically never be ignored, because some repositories are
unreadable without certain extensions or configurations.
The JavaHg API here mirrors the hg behavior for the `HGRCPATH` environment
variable- not setting it (or null here) uses the default config resolution, and
`""` (the default in JavaHg for some reason) disables default config resolution.
The repository initialization code in JavaHg also uses `""`, and there's no way
to alter that from the outside. But that appears to be harmless, so I'm
ignoring that for now. Note that this may only have been a problem on
non-Windows systems- setting an environment variable to empty on Windows
*unsets* the variable, even in the `_wputenv()` API, which is what we want.
After this, normal push/pull operations continue to use the global config. But
now imports from the SCM-Manager UI, the hooks run during a push, and any other
commands that are run through JavaHg will see a consistent configuration. LFS
(and maybe largefiles- I haven't tested it) repository imports now work. (I
wouldn't say "supported" yet because it doesn't pull the blobs. The largefiles
extension has a command for doing this, but the LFS extension doesn't. The
stopgap for this is to run `hg verify` to download the blobs, but that won't
work here with the way subrepos aren't nested as expected. I can work on a
command on the hg side to fill this gap.)
One final thing to note here- as I was testing, I initially got authentication
failures when trying to pull. It happened several times as I added more and
more logging, then disappeared, and I wasn't able to trigger it again as I
backed off the logging to get to this change alone. The auth info is written to
the repository's hgrc file, so a change to whether or not the global
configuration is processed is irrelevant. The next couple of commits will try
to improve the related code.
* Stop modifying and reusing the global JavaHg RepositoryConfiguration singleton
I can't point to a specific error, but modifying a singleton and reusing it is
a good way to get unexpected state. The extension adding won't collect
additional state (it's add-if-not-present), but the underlying
`java.util.HashMap` isn't threadsafe. Any differences in the environment map
when this is called would alter that state of anything else that still held a
configuration (and it also uses `HashMap`), and the pending changeset, encoding,
and HgBin settings would be overwritten for everybody outright.
There's only a default constructor for this class, so nothing to pass along in
the constructor. I don't *think* this was the cause of the random auth failures
mentioned in the previous commit, but it's easy to avoid these potential
problems.
* Avoid modifying the Mercurial repository config file to add config options
Modifying the config file is complicated and error prone, but there hasn't been
any other option aside from setting `HGRCPATH` to point to a standalone file.
Starting with Mercurial 7.0 (scheduled for March 2025), there's now a global
option to specify one or more additional config files on the command line,
without disabling the normal system and user level config file processing.[1]
Since I'm not sure what the minimum supported Mercurial is for this project,
this includes an extension that backports the same option if it is not present
in the version of Mercurial that is used, and does nothing if Mercurial natively
supports it. I tested back to Mercurial 6.0, which should be more than
sufficient- 6.1.4 (June 2022) was the last release to support Python 2 (which
has been EOL since Jan 2020), and the Python 3 support before that release was
considered experimental. It likely works in earlier versions, but there's a
definite minimum of 4.9 (Feb 2019), due to the `exthelper` module import.
Without the need to modify a possibly existing file and then restore it when
done, a bunch of code falls away, and the tests that supported it.
[1] https://repo.mercurial-scm.org/hg/rev/25b344f2aeef
* Clean up
---------
Co-authored-by: René Pfeuffer <rene.pfeuffer@cloudogu.com>
When a commit could not be found, JGit throws a MissingObjectException.
Because this extends IOException, the according catch block is executed
with a log with level ERROR and the message, that the repository could not
be opened. This is simply wrong.
With this, we catch the MissingObjectException explicitly and simply return null.
This performance optimization results from the fact that during namespace loading, read permissions are no longer verified for every repository, but only for those that would result in a new namespace. Simply put: If a namespace has already been discovered, no further repositories within that namespace are checked.
We need the new function `streamAll` for this, because the `getAllRepositories` would create all repositories before returning them, which would make it impossible to intervene in time using the filter parameter.
- Fix startup issue with unfinished export of deleted repository
- Add withIgnoreReadOnly setting
- Build store with ignore read only flag
- Add logging
With this, we updated the zipped test repositories in
scm-plugins/scm-hg-plugin/src/test/resources/sonia/scm/repository
with `hg debugupgraderepo --run` and a hg version 5.3.1
It lookes like the old format does not work with newer
version of hg (maybe since 6.x or 6.7 or something)
Squash commits of branch feature/diff_tree_extension:
- Add optional extension for diff file tree
- Make it a wrapper
- Make extension more complex
- Log change
