Replace key feedback with label instead of FeedbackMessage
Since it is a pest to get rid of a Wicket FeedbackMessage in an AJAX target,
change the code to use an extra label that can provide feedback if the
key could not be parsed or was empty.
authority: Fix null pointer crash for deleted users
When a user had a certificate, i.e. an entry in the Gitblit Authority
database, but the user was deleted from the Gitblit database, then the
Authority application crashes upon loading. This patch prevents the
crash. The deleted user is no longer shown in the Authority. But the
database entry still is kept. This should be improved to show deleted
users and give the possibility to delete them from the Authority's
database.
This fixes #1359
Since we now do not generate a DSA host key file anymore, but keep it in
the list of potential keys so that existing keys still work, it can
happen that the files for DSA (and Ed25519) are getting loaded but they
do not exist. This results in an error in the log.
So instead check if the file exists and only try to load files that
exist. This prevents from errors (which are none) being spammed in the
log.
Create new host keys, one with ECDSA and one with Ed25519 algorithms.
For the Ed25519 currently the EdDSA library from i2p is used. This
requires some quirks, compared to a modern BouncyCastle. But the SSHD
library used cannot use BouncyCastle yet for Ed25519.
No DSA key is generated anymore, but we still support existing ones.
The version 1.69 is chosen instead of 1.70, because the moxie build
would not download the jars, trying to download `...1.7.jar` instead.
Three class deprecations are fixed. `PEMWriter` and `X509Extension`
are replaced with their drop-in replacements `JcaPEMWriter` and
`Extension`. The `PasswordFinder` deprecation note says that "it is
no longer used". It also was never used in Gitblit's code, so it is
removed from the key par provider class.
Instead of adding another setting and having to explain how the new one
and the existing `requireClientCertificates` setting are interdependent,
let's use the existing setting and add new values.
It is changed from a boolean to a string, with the values `required`,
`optional` and `disabled`. To keep backward compatibility with the old
values, the `true` value is mapped to `required` and the `false` value
is mapped to `optional`.
This also removes the additional Set. What changes is that the order of
the repositories is kept, but the resulting list will have user and team
permissions intertwined.
fix: Fix StoredUserConfig not escaping control characters
The `StoredUserConfig` only escaped the escape character, i.e. backslash.
But it does not escape control characters like tab or newline. This
introduces a vulnerability where an attacker can create new entries
in their user account and create new accounts.
In addition, other characters are also not properly handled. Field values
with a comment character need to be quoted. This only happens for the
`#` character and only when the value starts with it. Also the quote
is note escaped in values.
This change completely rewrites the `escape` method of `StoredUserConfig`.
It takes care of properly escaping characters that need escaping for the
git configuration file format.
This fixes #1410
Te `StoredUserConfig` did not handle sections without a subsection.
When the subsection did not exist, i.e. was `null`, then the subsection
name would be set to the string "null". This is not how the config file
format works. It should create a `[SECTIONNAME]` entry instead.
This fix handles a `null` subsection correctly, by handling it as a
section without a subsection.
site: Old packages moved from Bintray to Cloudsmith
Since Bintray seized operation, the packages for versions 1.4.0 - 1.8.0
have been uploaded to Cloudsmith. Adjust the links appropriately.
An attribution is added to comply with Cloudsmith's OSS hosting policy.
To help with analysis, try to detect if the instance is running inside
a container. Some containers are detected, but this is probably not
exhaustive. At least a Docker container should be detectable.
Report in the runtime manager to the log if a container was detected.
pages: Fix repo-relative reference links in markdown docs
As with explicit links, also for reference links in markdown documents
which point to repository-relative files the links are broken. They do
not take the path to the repository into account.
This fix is related to commit b23269 which fixed issue #1358
for explicit links.
bug: Fix double encoding links in Markdown/Wiki pages
When parsing Markdown or Wiki pages, links get URL encoded. This happened
twice for links to other documents. Once explicitly and once by Wicket
when it creates a `urlFor` the page. That results in multi-byte
characters getting percent escaped, and then the percent character again
getting percent escaped.
The explicit encoding looks like a forgotten left over, so it gets
removed from the code. The Wicket encoding is smarter anyways, knowing
what is path and what is parameter.
This fixes #864.
Update the URL to the Gitblit docker image to point to the
`gitblit/gitblit` repository, which I now consider the official
Docker repository.
Also link to the Docker images on the main page.
Although it seems strange to have a RefModel with a referenced object
but a null Ref, Gitblit uses such RefModels for instance in
JGitUtils.getNotesOnCommit().
Be careful to do something sensible when that Ref is null.
JGit commit objects are a recursive data structure; they have links to
their parent commits. Serializing a JGit commit will try to recursively
serialize all reachable ancestors as faras they have been loaded. If
that ancestor chain is too long, a StackOverflowError is thrown during
Wicket's page serialization if a page has a reference to sucha JGit
commit.
Fixed by making sure that pages o not contain references to JGit
commits. Use the (existing) wrapper object RepositoryCommit instead.
* RepositoryCommit has a transient reference to the JGit commit and
reads the commit from the repository upon de-serialization.
* RefModel is a similar case (JGit tags/branches may also have links
to the commits they point to). Solved a bit differently by making it
a pure data object by transferring the interesting data from the JGit
object in the constructor.
* Change DataViews instantiated with RevCommit to use RepositoryCommit
instead.
* Change inner anonymous DataViews to ensure they do not have a
synthesized field referencing the "allRefs" map. Such a synthesized
field would also get serialized, and then serialize JGit commits
again.
Finally, remove non-transient logger instances in Wicket classes. Those
might lead to NotSerializableException.
These StackOverflowErrors have been reported in several places since
2014:
* https://groups.google.com/forum/#!topic/gitblit/GH1d8WSlR6Q
* https://bugs.chromium.org/p/gerrit/issues/detail?id=3316
* https://groups.google.com/d/msg/repo-discuss/Kcl0JIGNiGk/0DjH4mO8hA8J
* https://groups.google.com/d/msg/repo-discuss/0_P6A3fjTec/2kcpVPIUAQAJ
* https://github.com/gitblit/gitblit/issues/1011
* https://github.com/tomaswolf/gerrit-gitblit-plugin/issues/21
So far links to raw view were not encoded. The browser did some encoding
of spaces on its own, which the servlet would unescape, since it uses
the `HttpServletRequest.getPathInfo` method. That decodes the path
before returning it.
A problem arises when a bracket is in the file (or folder) name. The
brackets are the characters that are not allowed in the path, according
to the `URI.parse` method. (Which is a bit harsh, because brackets
actually are only reserved for the host part since IPv6.) That means
that the decoding fails when a bracket character is encountered.
This went unnoticed since the failed decoding will return the path
as it got it. But once there is a space in the file name, which the
browser helpfully encoded for us, the failed decoding will now leave the
encoded space in there. And that will result in a path that does not
exist, e.g. `file%20[a]`.
To be on the safe side, we simply encode the path in the links that we
generate, so that it complies with the rules that are used in `getPathInfo`.
This fixes #1375.
TimeUtils: Change daysAgo to calculate difference in calendar days
The `daysAgo` method seemed to want to normalize on a calendar day? I
can't really tell what it was trying to do, but the problem is that it
does not take into account any time shift due to time zones so it never
really worked outside of GMT.
So instead a new `calendarDaysAgo` method is added (because I am unsure
on what the `daysAgo` method is trying to do. It can probably be removed).
The new method cleanly calculates difference in calendar days because it
normalizes the two given time stamps on the same time zone.
The `timeAgo` method now used the new method. This fixes #1248.
TimeUtils: Move unit test to same package as TimeUtils is
For some reason the `TimeUtilsTest` class is, like almost all tests, in
the `com.gitblit.tests` package. But this way all methods in classes
which we might predominately need for tests have to be public.
So move the unit test class `TimeUtilsTest` to the same package as the
class it is testing, i.e. `com.gitblit.utils.TimeUtils`.
This way we ca set the new added methods which get the current time
passed in to be at least not public.
Add tests for `timeAgo` to analyse issue #1248.
The tests are dependent on when they run as they time functions use the
current date and time. To make them testable in a reproducible way, we
need the ability to pass in what we think is "now". So add overloaded
methods that take a `now` parameter so that we can pass in the current
time.
Some property keys had typos.
There is a `gb.ticketStatus` and a `gb.ticketState`. Neither is used
anywhere in the code, but only the former is defined in the default file.
So only use `gb.ticketStatus`.