We just safe-guarded `.git` against NTFS Alternate Data Stream-related
attack vectors, and now it is time to do the same for `.gitmodules`.
Note: In the added regression test, we refrain from verifying all kinds
of variations between short names and NTFS Alternate Data Streams: as
the new code disallows _all_ Alternate Data Streams of `.gitmodules`, it
is enough to test one in order to know that all of them are guarded
against.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
We need to be careful to follow proper quoting rules. For example, if an
argument contains spaces, we have to quote them. Double-quotes need to
be escaped. Backslashes need to be escaped, but only if they are
followed by a double-quote character.
We need to be _extra_ careful to consider the case where an argument
ends in a backslash _and_ needs to be quoted: in this case, we append a
double-quote character, i.e. the backslash now has to be escaped!
The current code, however, fails to recognize that, and therefore can
turn an argument that ends in a single backslash into a quoted argument
that now ends in an escaped double-quote character. This allows
subsequent command-line parameters to be split and part of them being
mistaken for command-line options, e.g. through a maliciously-crafted
submodule URL during a recursive clone.
Technically, we would not need to quote _all_ arguments which end in a
backslash _unless_ the argument needs to be quoted anyway. For example,
`test\` would not need to be quoted, while `test \` would need to be.
To keep the code simple, however, and therefore easier to reason about
and ensure its correctness, we now _always_ quote an argument that ends
in a backslash.
This addresses CVE-2019-1350.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Previously, this function was written without focusing on speed,
intending to make reviewing the code as easy as possible, to avoid any
bugs in this critical code.
Turns out: we can do much better on both accounts. With this patch, we
make it as fast as this developer can make it go:
- We avoid the call to `is_dir_sep()` and make all the character
comparisons explicit.
- We avoid the cost of calling `strncasecmp()` and unroll the test for
`.git` and `git~1`, not even using `tolower()` because it is faster to
compare against two constant values.
- We look for `.git` and `.git~1` first thing, and return early if not
found.
- We also avoid calling a separate function for detecting chains of
spaces and periods.
Each of these improvements has a noticeable impact on the speed of
`is_ntfs_dotgit()`.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Probably inspired by HFS' resource streams, NTFS supports "Alternate
Data Streams": by appending `:<stream-name>` to the file name,
information in addition to the file contents can be written and read,
information that is copied together with the file (unless copied to a
non-NTFS location).
These Alternate Data Streams are typically used for things like marking
an executable as having just been downloaded from the internet (and
hence not necessarily being trustworthy).
In addition to a stream name, a stream type can be appended, like so:
`:<stream-name>:<stream-type>`. Unless specified, the default stream
type is `$DATA` for files and `$INDEX_ALLOCATION` for directories. In
other words, `.git::$INDEX_ALLOCATION` is a valid way to reference the
`.git` directory!
In our work in Git v2.2.1 to protect Git on NTFS drives under
`core.protectNTFS`, we focused exclusively on NTFS short names, unaware
of the fact that NTFS Alternate Data Streams offer a similar attack
vector.
Let's fix this.
Seeing as it is better to be safe than sorry, we simply disallow paths
referring to *any* NTFS Alternate Data Stream of `.git`, not just
`::$INDEX_ALLOCATION`. This also simplifies the implementation.
This closes CVE-2019-1352.
Further reading about NTFS Alternate Data Streams:
https://docs.microsoft.com/en-us/openspecs/windows_protocols/ms-fscc/c54dec26-1551-4d3a-a0ea-4fa40f848eb3
Reported-by: Nicolas Joly <Nicolas.Joly@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The config setting `core.protectNTFS` is specifically designed to work
not only on Windows, but anywhere, to allow for repositories hosted on,
say, Linux servers to be protected against NTFS-specific attack vectors.
As a consequence, `is_ntfs_dotgit()` manually splits backslash-separated
paths (but does not do the same for paths separated by forward slashes),
under the assumption that the backslash might not be a valid directory
separator on the _current_ Operating System.
However, the two callers, `verify_path()` and `fsck_tree()`, are
supposed to feed only individual path segments to the `is_ntfs_dotgit()`
function.
This causes a lot of duplicate scanning (and very inefficient scanning,
too, as the inner loop of `is_ntfs_dotgit()` was optimized for
readability rather than for speed.
Let's simplify the design of `is_ntfs_dotgit()` by putting the burden of
splitting the paths by backslashes as directory separators on the
callers of said function.
Consequently, the `verify_path()` function, which already splits the
path by directory separators, now treats backslashes as directory
separators _explicitly_ when `core.protectNTFS` is turned on, even on
platforms where the backslash is _not_ a directory separator.
Note that we have to repeat some code in `verify_path()`: if the
backslash is not a directory separator on the current Operating System,
we want to allow file names like `\`, but we _do_ want to disallow paths
that are clearly intended to cause harm when the repository is cloned on
Windows.
The `fsck_tree()` function (the other caller of `is_ntfs_dotgit()`) now
needs to look for backslashes in tree entries' names specifically when
`core.protectNTFS` is turned on. While it would be tempting to
completely disallow backslashes in that case (much like `fsck` reports
names containing forward slashes as "full paths"), this would be
overzealous: when `core.protectNTFS` is turned on in a non-Windows
setup, backslashes are perfectly valid characters in file names while we
_still_ want to disallow tree entries that are clearly designed to
exploit NTFS-specific behavior.
This simplification will make subsequent changes easier to implement,
such as turning `core.protectNTFS` on by default (not only on Windows)
or protecting against attack vectors involving NTFS Alternate Data
Streams.
Incidentally, this change allows for catching malicious repositories
that contain tree entries of the form `dir\.gitmodules` already on the
server side rather than only on the client side (and previously only on
Windows): in contrast to `is_ntfs_dotgit()`, the
`is_ntfs_dotgitmodules()` function already expects the caller to split
the paths by directory separators.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
In preparation to flipping the default on `core.protectNTFS`, let's have
some way to measure the speed impact of this config setting reliably
(and for comparison, the `core.protectHFS` config setting).
For now, this is a manual performance benchmark:
./t/helper/test-path-utils protect_ntfs_hfs [arguments...]
where the arguments are an optional number of file names to test with,
optionally followed by minimum and maximum length of the random file
names. The default values are one million, 3 and 20, respectively.
Just like `sqrti()` in `bisect.c`, we introduce a very simple function
to approximation the square root of a given value, in order to avoid
having to introduce the first user of `<math.h>` in Git's source code.
Note: this is _not_ implemented as a Unix shell script in t/perf/
because we really care about _very_ precise timings here, and Unix shell
scripts are simply unsuited for precise and consistent benchmarking.
Signed-off-by: Garima Singh <garima.singh@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
This patch series fixes an issue where Git could formerly have been
tricked into creating a `.git` file with an unexpected (and therefore
unprotected) NTFS short name.
Incidentally, it also fixes an issue where a tree entry containing a
backslash could be tricked into following a symbolic link, i.e. Git
could be tricked into writing files outside the worktree.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The `--export-marks` option of `git fast-import` is exposed also via the
in-stream command `feature export-marks=...` and it allows overwriting
arbitrary paths.
This topic branch prevents the in-stream version, to prevent arbitrary
file accesses by `git fast-import` streams coming from untrusted sources
(e.g. in remote helpers that are based on `git fast-import`).
This fixes CVE-2019-1348.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Previously, this function was completely undocumented. It is worth,
though, to explain what is going on, as it is not really obvious at all.
Suggested-by: Garima Singh <garima.singh@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The backslash character is not a valid part of a file name on Windows.
Hence it is dangerous to allow writing files that were unpacked from
tree objects, when the stored file name contains a backslash character:
it will be misinterpreted as directory separator.
This not only causes ambiguity when a tree contains a blob `a\b` and a
tree `a` that contains a blob `b`, but it also can be used as part of an
attack vector to side-step the careful protections against writing into
the `.git/` directory during a clone of a maliciously-crafted
repository.
Let's prevent that, addressing CVE-2019-1354.
Note: we guard against backslash characters in tree objects' file names
_only_ on Windows (because on other platforms, even on those where NTFS
volumes can be mounted, the backslash character is _not_ a directory
separator), and _only_ when `core.protectNTFS = true` (because users
might need to generate tree objects for other platforms, of course
without touching the worktree, e.g. using `git update-index
--cacheinfo`).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
In addition to preventing `.git` from being tracked by Git, on Windows
we also have to prevent `git~1` from being tracked, as the default NTFS
short name (also known as the "8.3 filename") for the file name `.git`
is `git~1`, otherwise it would be possible for malicious repositories to
write directly into the `.git/` directory, e.g. a `post-checkout` hook
that would then be executed _during_ a recursive clone.
When we implemented appropriate protections in 2b4c6efc82 (read-cache:
optionally disallow NTFS .git variants, 2014-12-16), we had analyzed
carefully that the `.git` directory or file would be guaranteed to be
the first directory entry to be written. Otherwise it would be possible
e.g. for a file named `..git` to be assigned the short name `git~1` and
subsequently, the short name generated for `.git` would be `git~2`. Or
`git~3`. Or even `~9999999` (for a detailed explanation of the lengths
we have to go to protect `.gitmodules`, see the commit message of
e7cb0b4455 (is_ntfs_dotgit: match other .git files, 2018-05-11)).
However, by exploiting two issues (that will be addressed in a related
patch series close by), it is currently possible to clone a submodule
into a non-empty directory:
- On Windows, file names cannot end in a space or a period (for
historical reasons: the period separating the base name from the file
extension was not actually written to disk, and the base name/file
extension was space-padded to the full 8/3 characters, respectively).
Helpfully, when creating a directory under the name, say, `sub.`, that
trailing period is trimmed automatically and the actual name on disk
is `sub`.
This means that while Git thinks that the submodule names `sub` and
`sub.` are different, they both access `.git/modules/sub/`.
- While the backslash character is a valid file name character on Linux,
it is not so on Windows. As Git tries to be cross-platform, it
therefore allows backslash characters in the file names stored in tree
objects.
Which means that it is totally possible that a submodule `c` sits next
to a file `c\..git`, and on Windows, during recursive clone a file
called `..git` will be written into `c/`, of course _before_ the
submodule is cloned.
Note that the actual exploit is not quite as simple as having a
submodule `c` next to a file `c\..git`, as we have to make sure that the
directory `.git/modules/b` already exists when the submodule is checked
out, otherwise a different code path is taken in `module_clone()` that
does _not_ allow a non-empty submodule directory to exist already.
Even if we will address both issues nearby (the next commit will
disallow backslash characters in tree entries' file names on Windows,
and another patch will disallow creating directories/files with trailing
spaces or periods), it is a wise idea to defend in depth against this
sort of attack vector: when submodules are cloned recursively, we now
_require_ the directory to be empty, addressing CVE-2019-1349.
Note: the code path we patch is shared with the code path of `git
submodule update --init`, which must not expect, in general, that the
directory is empty. Hence we have to introduce the new option
`--force-init` and hand it all the way down from `git submodule` to the
actual `git submodule--helper` process that performs the initial clone.
Reported-by: Nicolas Joly <Nicolas.Joly@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
As with export-marks in the previous commit, import-marks can access the
filesystem. This is significantly less dangerous than export-marks
because it only involves reading from arbitrary paths, rather than
writing them. However, it could still be surprising and have security
implications (e.g., exfiltrating data from a service that accepts
fast-import streams).
Let's lump it (and its "if-exists" counterpart) in with export-marks,
and enable the in-stream version only if --allow-unsafe-features is set.
Signed-off-by: Jeff King <peff@peff.net>
The fast-import stream command "feature export-marks=<path>" lets the
stream write marks to an arbitrary path. This may be surprising if you
are running fast-import against an untrusted input (which otherwise
cannot do anything except update Git objects and refs).
Let's disallow the use of this feature by default, and provide a
command-line option to re-enable it (you can always just use the
command-line --export-marks as well, but the in-stream version provides
an easy way for exporters to control the process).
This is a backwards-incompatible change, since the default is flipping
to the new, safer behavior. However, since the main users of the
in-stream versions would be import/export-based remote helpers, and
since we trust remote helpers already (which are already running
arbitrary code), we'll pass the new option by default when reading a
remote helper's stream. This should minimize the impact.
Note that the implementation isn't totally simple, as we have to work
around the fact that fast-import doesn't parse its command-line options
until after it has read any "feature" lines from the stream. This is how
it lets command-line options override in-stream. But in our case, it's
important to parse the new --allow-unsafe-features first.
There are three options for resolving this:
1. Do a separate "early" pass over the options. This is easy for us to
do because there are no command-line options that allow the
"unstuck" form (so there's no chance of us mistaking an argument
for an option), though it does introduce a risk of incorrect
parsing later (e.g,. if we convert to parse-options).
2. Move the option parsing phase back to the start of the program, but
teach the stream-reading code never to override an existing value.
This is tricky, because stream "feature" lines override each other
(meaning we'd have to start tracking the source for every option).
3. Accept that we might parse a "feature export-marks" line that is
forbidden, as long we don't _act_ on it until after we've parsed
the command line options.
This would, in fact, work with the current code, but only because
the previous patch fixed the export-marks parser to avoid touching
the filesystem.
So while it works, it does carry risk of somebody getting it wrong
in the future in a rather subtle and unsafe way.
I've gone with option (1) here as simple, safe, and unlikely to cause
regressions.
This fixes CVE-2019-1348.
Signed-off-by: Jeff King <peff@peff.net>
When we parse the --export-marks option, we don't immediately open the
file, but we do create any leading directories. This can be especially
confusing when a command-line option overrides an in-stream one, in
which case we'd create the leading directory for the in-stream file,
even though we never actually write the file.
Let's instead create the directories just before opening the file, which
means we'll create only useful directories. Note that this could change
the handling of relative paths if we chdir() in between, but we don't
actually do so; the only permanent chdir is from setup_git_directory()
which runs before either code path (potentially we should take the
pre-setup dir into account to avoid surprising the user, but that's an
orthogonal change).
The test just adapts the existing "override" test to use paths with
leading directories. This checks both that the correct directory is
created (which worked before but was not tested), and that the
overridden one is not (our new fix here).
While we're here, let's also check the error result of
safe_create_leading_directories(). We'd presumably notice any failure
immediately after when we try to open the file itself, but we can give a
more specific error message in this case.
Signed-off-by: Jeff King <peff@peff.net>
When asked to import marks from "subdir/file.marks", we create the
leading directory "subdir" if it doesn't exist. This makes no sense for
importing marks, where we only ever open the path for reading.
Most of the time this would be a noop, since if the marks file exists,
then the leading directories exist, too. But if it doesn't (e.g.,
because --import-marks-if-exists was used), then we'd create the useless
directory.
This dates back to 580d5f83e7 (fast-import: always create marks_file
directories, 2010-03-29). Even then it was useless, so it seems to have
been added in error alongside the --export-marks case (which _is_
helpful).
Signed-off-by: Jeff King <peff@peff.net>
We parse options like "--max-pack-size=" using skip_prefix(), which
makes sense to get at the bytes after the "=". However, we also parse
"--quiet" and "--stats" with skip_prefix(), which allows things like
"--quiet-nonsense" to behave like "--quiet".
This was a mistaken conversion in 0f6927c229 (fast-import: put option
parsing code in separate functions, 2009-12-04). Let's tighten this to
an exact match, which was the original intent.
Signed-off-by: Jeff King <peff@peff.net>
Our tests confirm that providing two "import-marks" options in a
fast-import stream is an error. However, the invoked command would fail
even without covering this case, because the marks files themselves do
not actually exist. Let's create the files to make sure we fail for the
right reason (we actually do, because the option parsing happens before
we open anything, but this future-proofs our test).
Signed-off-by: Jeff King <peff@peff.net>
Finishing touches to the recent update to the build procedure for
the documentation.
* bc/doc-use-docbook-5:
manpage-bold-literal.xsl: match for namespaced "d:literal" in template
* https://github.com/prati0100/git-gui:
git-gui: improve Japanese translation
git-gui: add a readme
git-gui: support for diff3 conflict style
git-gui: use existing interface to query a path's attribute
git-gui (Windows): use git-bash.exe if it is available
treewide: correct several "up-to-date" to "up to date"
Fix build with core.autocrlf=true
As per Wikipedia, "In current technical usage, for one to state that a
feature is deprecated is merely a recommendation against using it." It
is thus contradictory to claim that something is not "officially
deprecated" and then to immediately state that we are both discouraging
its use and pointing people elsewhere.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We recently regressed our rendering with Asciidoctor of "literal"
elements in our manpages, i.e, stuff we have placed within `backticks`
in order to render as monospace. In particular, we lost the bold
rendering of such literal text.
The culprit is f6461b82b9 ("Documentation: fix build with Asciidoctor 2",
2019-09-15), where we switched from DocBook 4.5 to DocBook 5 with
Asciidoctor. As part of the switch, we started using the namespaced
DocBook XSLT stylesheets rather than the non-namespaced ones. (See
f6461b82b9 for more details on why we changed to the namespaced ones.)
The bold literals are implemented as an XSLT snippet <xsl:template
match="literal">...</xsl:template>. Now that we use namespaces, this
doesn't pick up our literals like it used to.
Match for "d:literal" in addition to just "literal", after defining the
d namespace. ("d" is what
http://docbook.sourceforge.net/release/xsl-ns/current/manpages/docbook.xsl
uses.) Note that we need to keep matching without the namespace for
AsciiDoc.
This boldness was introduced by 5121a6d993 ("Documentation: option to
render literal text as bold for manpages", 2009-03-27) and made the
default in 5945717009 ("Documentation: bold literals in man",
2016-05-31).
One reason this was not caught in review is that our doc-diff tool diffs
without any boldness, i.e., it "only" compares text. As pointed out by
Peff in review of this patch, one can use `MAN_KEEP_FORMATTING=1
./doc-diff <...>`
This has been optically tested with AsciiDoc 8.6.10, Asciidoctor 1.5.5
and Asciidoctor 2.0.10. I've also verified that doc-diff produces the
empty output for all three programs, as expected, and that with the
MAN_KEEP_FORMATTING trick, AsciiDoc yields no diff, whereas with
Asciidoctor, we get bold literals, just like we want.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Acked-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix a rather old bug in gitweb---incremental blame output in
javascript actions mode never worked.
* rl/gitweb-blame-prev-fix:
gitweb: correctly store previous rev in javascript-actions mode
The comments for the staging/unstaging test did not accurately
describe the scenario being tested. It is not essential that
the test files being staged/unstaged appear at the end of the
index. All that is required is that the test files are not
flagged with CE_FSMONITOR_VALID and have a position in the
index greater than the number of entries in the index after
unstaging.
The comment for this test has been updated to be more
accurate with respect to the scenario that's being tested.
Signed-off-by: William Baker <William.Baker@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
-----BEGIN PGP SIGNATURE-----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=zfW8
-----END PGP SIGNATURE-----
Merge tag 'v2.24.0-rc1' of github.com:git/git into master
Git 2.24-rc1
* tag 'v2.24.0-rc1' of github.com:git/git:
Git 2.24-rc1
repo-settings: read an int for index.version
ci: fix GCC install in the Travis CI GCC OSX job
Eleventh batch
ci(osx): use new location of the `perforce` cask
t7419: change test_must_fail to ! for grep
t4014: make output-directory tests self-contained
ci(visual-studio): actually run the tests in parallel
ci(visual-studio): use strict compile flags, and optimization
userdiff: fix some corner cases in dts regex
test-progress: fix test failures on big-endian systems
completion: clarify installation instruction for zsh
grep: avoid leak of chartables in PCRE2
grep: make PCRE2 aware of custom allocator
grep: make PCRE1 aware of custom allocator
remote-curl: pass on atomic capability to remote side
diff-highlight: fix a whitespace nit
fsmonitor: don't fill bitmap with entries to be removed
When this function is passed a path with a trailing slash, it runs right
over the end of that path.
Let's fix this.
Co-authored-by: Alexandr Miloslavskiy <alexandr.miloslavskiy@syntevo.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Without this change, the setting
$feature{'javascript-actions'}{'default'} = [1];
in gitweb.conf breaks gitweb's blame page: clicking on line numbers
displayed in the second column on the page has no effect.
For comparison, with javascript-actions disabled, clicking on line
numbers loads the previous version of the line.
Addresses https://bugs.debian.org/741883.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Robert Luberda <robert@debian.org>
Acked-by: Jakub Narębski <jnareb@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous commit includes a failing test for an issue around
fetch.writeCommitGraph and fetching in a repo with a submodule. Here, we
fix that bug and set the test to "test_expect_success".
The problem arises with this set of commands when the remote repo at
<url> has a submodule. Note that --recurse-submodules is not needed to
demonstrate the bug.
$ git clone <url> test
$ cd test
$ git -c fetch.writeCommitGraph=true fetch origin
Computing commit graph generation numbers: 100% (12/12), done.
BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
Aborted (core dumped)
As an initial fix, I converted the code in builtin/fetch.c that calls
write_commit_graph_reachable() to instead launch a "git commit-graph
write --reachable --split" process. That code worked, but is not how we
want the feature to work long-term.
That test did demonstrate that the issue must be something to do with
internal state of the 'git fetch' process.
The write_commit_graph() method in commit-graph.c ensures the commits we
plan to write are "closed under reachability" using close_reachable().
This method walks from the input commits, and uses the UNINTERESTING
flag to mark which commits have already been visited. This allows the
walk to take O(N) time, where N is the number of commits, instead of
O(P) time, where P is the number of paths. (The number of paths can be
exponential in the number of commits.)
However, the UNINTERESTING flag is used in lots of places in the
codebase. This flag usually means some barrier to stop a commit walk,
such as in revision-walking to compare histories. It is not often
cleared after the walk completes because the starting points of those
walks do not have the UNINTERESTING flag, and clear_commit_marks() would
stop immediately.
This is happening during a 'git fetch' call with a remote. The fetch
negotiation is comparing the remote refs with the local refs and marking
some commits as UNINTERESTING.
I tested running clear_commit_marks_many() to clear the UNINTERESTING
flag inside close_reachable(), but the tips did not have the flag, so
that did nothing.
It turns out that the calculate_changed_submodule_paths() method is at
fault. Thanks, Peff, for pointing out this detail! More specifically,
for each submodule, the collect_changed_submodules() runs a revision
walk to essentially do file-history on the list of submodules. That
revision walk marks commits UNININTERESTING if they are simplified away
by not changing the submodule.
Instead, I finally arrived on the conclusion that I should use a flag
that is not used in any other part of the code. In commit-reach.c, a
number of flags were defined for commit walk algorithms. The REACHABLE
flag seemed like it made the most sense, and it seems it was not
actually used in the file. The REACHABLE flag was used in early versions
of commit-reach.c, but was removed by 4fbcca4 (commit-reach: make
can_all_from_reach... linear, 2018-07-20).
Add the REACHABLE flag to commit-graph.c and use it instead of
UNINTERESTING in close_reachable(). This fixes the bug in manual
testing.
Reported-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Helped-by: Jeff King <peff@peff.net>
Helped-by: Szeder Gábor <szeder.dev@gmail.com>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
While dogfooding, Johannes found a bug in the fetch.writeCommitGraph
config behavior. His example initially happened during a clone with
--recurse-submodules, we found that this happens with the first fetch
after cloning a repository that contains a submodule:
$ git clone <url> test
$ cd test
$ git -c fetch.writeCommitGraph=true fetch origin
Computing commit graph generation numbers: 100% (12/12), done.
BUG: commit-graph.c:886: missing parent <hash1> for commit <hash2>
Aborted (core dumped)
In the repo I had cloned, there were really 60 commits to scan, but
only 12 were in the list to write when calling
compute_generation_numbers(). A commit in the list expects to see a
parent, but that parent is not in the list.
A follow-up will fix the bug, but first we create a test that
demonstrates the problem. This test must be careful about an existing
commit-graph file, since GIT_TEST_COMMIT_GRAPH=1 will cause the repo we
are cloning to already have one. This then prevents the incremtnal
commit-graph write during the first 'git fetch'.
Helped-by: Jeff King <peff@peff.net>
Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Helped-by: Szeder Gábor <szeder.dev@gmail.com>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Johannes Schindelin