git_rebase_finish relies on head_detached being set, but
rebase_init_merge was only setting it when branch->ref_name was unset.
But branch->ref_name would be set to "HEAD" in the case of detached
HEAD being either implicitly (NULL) or explicitly passed to
git_rebase_init.
Introduce `git_thread_exit`, which will allow threads to terminate at an
arbitrary time, returning a `void *`. On Windows, this means that we
need to store the current `git_thread` in TLS, so that we can set its
`return` value when terminating.
We cannot simply use `ExitThread`, since Win32 returns `DWORD`s from
threads; we return `void *`.
`giterr_set()` is used when it is required to format a string, and since
we don't really require it for this case, it is better to stick to
`giterr_set_str()`.
This also suppresses a warning(-Wformat-security) raised by the compiler.
Signed-off-by: Pranit Bauva <pranit.bauva@gmail.com>
The `CURLINFO_LASTSOCKET` information has been deprecated since
curl version 7.45.0 as it may result in an overflow in the
returned socket on certain systems, most importantly on 64 bit
Windows. Instead, a new call `CURLINFO_ACTIVESOCKET` has been
added which instead returns a `curl_socket_t`, which is always
sufficiently long to store a socket.
As we need to provide backwards compatibility with curl versions
smaller than 7.45.0, alias CURLINFO_ACTIVESOCKET to
CURLINFO_LASTSOCKET on platforms without CURLINFO_ACTIVESOCKET.
We look at whether we're trying to replace a blob with a tree during the
update phase, but we fail to look at whether we've just inserted a blob
where we're now trying to insert a tree.
Update the check to look at both places. The test for this was
previously succeeding due to the bu where we did not look at the sorted
output.
On Windows we can find locked files even when reading a reference or the
packed-refs file. Bubble up the error in this case as well to allow
callers on Windows to retry more intelligently.
It does not help us to check whether the file exists before trying to
unlink it since it might be gone by the time unlink is called.
Instead try to remove it and handle the resulting error if it did not
exist.
Checking the size before we open the file descriptor can lead to the
file being replaced from under us when renames aren't quite atomic, so
we can end up reading too little of the file, leading to us thinking the
file is corrupted.
There might be a few threads or processes working with references
concurrently, so fortify the code to ignore errors which come from
concurrent access which do not stop us from continuing the work.
This includes ignoring an unlinking error. Either someone else removed
it or we leave the file around. In the former case the job is done, and
in the latter case, the ref is still in a valid state.
We need to save the errno, lest we clobber it in the giterr_set()
call. Also add code for reporting that a path component is missing,
which is a distinct failure mode.
In order not to undo concurrent modifications to references, we must
make sure that we only delete a loose reference if it still has the same
value as when we packed it.
This means we need to lock it and then compare the value with the one we
put in the packed file.
When trying to find a discovery, we walk up the directory
structure checking if there is a ".git" file or directory and, if
so, check its validity. But in the case that we've got a ".git"
file, we do not want to unconditionally assume that the file is
in fact a ".git" file and treat it as such, as we would error out
if it is not.
Fix the issue by only treating a file as a gitlink file if it
ends with "/.git". This allows users of the function to discover
a repository by handing in any path contained inside of a git
repository.
The existing code would set a namespace of "" (empty string) with
GIT_NAMESPACE unset. In a repository where refs/heads/namespaces/
exists, that can produce incorrect results. Detect that case and avoid
setting the namespace at all.
Since that makes the last assignment to error conditional, and the
previous assignment can potentially get GIT_ENOTFOUND, set error to 0
explicitly to prevent the call from incorrectly failing with
GIT_ENOTFOUND.
We're recently trying to upgrade to the current master of libgit2
in Cargo but we're unfortunately hitting a segfault in one of our
tests. This particular test is just a small smoke test that https
works (e.g. it's configured in libgit2). It attempts to clone
from a URL which simply immediately drops connections after
they're accepted (e.g. terminate abnormally). We expect to see a
standard error from libgit2 but unfortunately we're seeing a
segfault.
This segfault is happening inside of the `wait_for` function of
`curl_stream.c` at the line `FD_SET(fd, &errfd)` because `fd` is
-1. This ends up doing an out-of-bounds array access that faults
the program. I tracked back to where this -1 came from to the
line here (returned by `CURLINFO_LASTSOCKET`) and added a check
to return an error.
The code correctly detects that forced creation of a branch on a
nonbare repo should not be able to overwrite a branch which is
the HEAD reference. But there's no reason to prevent this on
a bare repo, and in fact, git allows this. I.e.,
git branch -f master new_sha
works on a bare repo with HEAD set to master. This change fixes
that problem, and updates tests so that, for this case, both the
bare and nonbare cases are checked for correct behavior.
We need to include the initialisation and construction functions in all
backend, so we include this header when building against SecureTransport
and WinHTTP as well.
In `pack_entry_find_offset`, we try to find the offset of a
certain object in the pack file. To do so, we first assert if the
packfile has already been opened and open it if not. Opening the
packfile is guarded with a mutex, so concurrent access to this is
in fact safe.
What is not thread-safe though is our calculation of offsets
inside the packfile. Assume two threads calling
`pack_entry_find_offset` at the same time. We first calculate the
offset and index location and only then determine if the pack has
already been opened. If so, we re-calculate the offset and index
address.
Now the case for two threads: thread 1 first calculates the
addresses and is subsequently suspended. The second thread will
now call `pack_index_open` and initialize the pack file,
calculating its addresses correctly. When the first thread is
resumed now, he'll see that the pack file has already been
initialized and will happily proceed with the addresses it has
already calculated before the check. As the pack file was not
initialized before, these addresses are bogus.
Fix the issue by only calculating the addresses after having
checked if the pack file is open.
When trying to receive packets from the remote, we loop until
either an error distinct to `GIT_EBUFS` occurs or until we
successfully parsed the packet. This does not honor the case
where we are looping over an already closed socket which has no
more data, leaving us in an infinite loop if we got a bogus
packet size or if the remote hang up.
Fix the issue by returning `GIT_EEOF` when we cannot read data
from the socket anymore.
When reading a server's reference announcements via the smart
protocol, we expect the server to send multiple flushes before
the protocol is finished. If we fail to receive new data from the
socket, we will only return an end of stream error if we have not
seen any flush yet.
This logic is flawed in that we may run into an infinite loop
when receiving a server's reference announcement with a bogus
flush packet. E.g. assume the last flushing package is changed to
not be '0000' but instead any other value. In this case, we will
still await one more flush package and ignore the fact that we
are not receiving any data from the socket, causing an infinite
loop.
Fix the issue by always returning `GIT_EEOF` if the socket
indicates an end of stream.
When threading is not enabled for libgit2, we keep global state
in a simple static variable. When libgit2 is shut down, we clean
up the global state by freeing the global state's dynamically
allocated memory. When libgit2 is built with threading, we
additionally free the thread-local storage and thus completely
remove the global state. In a non-threaded build, though, we
simply leave the global state as-is, which may result in an error
upon reinitializing libgit2.
Fix the issue by zeroing out the variable on a shutdown, thus
returning it to its initial state.
When trying to initialize and tear down global data structures
from different threads at once with `git_libgit2_init` and
`git_libgit2_shutdown`, we race around initializing data. While
we use `pthread_once` to assert that we only initilize data a
single time, we actually reset the `pthread_once_t` on the last
call to `git_libgit2_shutdown`. As resetting this variable is not
synchronized with other threads trying to access it, this is
actually racy when one thread tries to do a complete shutdown of
libgit2 while another thread tries to initialize it.
Fix the issue by creating a mutex which synchronizes `init_once`
and the library shutdown.
The `git_pqueue` struct allows being fixed in its total number of
entries. In this case, we simply throw away items that are
inserted into the priority queue by examining wether the new item
to be inserted has a higher priority than the previous smallest
one.
This feature somewhat contradicts our pqueue implementation in
that it is allowed to not have a comparison function. In fact, we
also fail to check if the comparison function is actually set in
the case where we add a new item into a fully filled fixed-size
pqueue.
As we cannot determine which item is the smallest item in absence
of a comparison function, we fix the `NULL` pointer dereference
by simply dropping all new items which are about to be inserted
into a full fixed-size pqueue.
Remove overriding the `checkout_strategy` for `update_options` when
performing an update on a submodule. Users should be specifying the
correct checkout strategy in
`update_options.checkout_opts.checkout_strategy`.
When parsing a commit, we will treat all bytes left after parsing
the headers as the commit message. When no bytes are left, we
leave the commit's message uninitialized. While uncommon to have
a commit without message, this is the right behavior as Git
unfortunately allows for empty commit messages.
Given that this scenario is so uncommon, most programs acting on
the commit message will never check if the message is actually
set, which may lead to errors. To work around the error and not
lay the burden of checking for empty commit messages to the
developer, initialize the commit message with an empty string
when no commit message is given.
When parsing tree entries from raw object data, we do not verify
that the tree entry actually has a filename as well as a valid
object ID. Fix this by asserting that the filename length is
non-zero as well as asserting that there are at least
`GIT_OID_RAWSZ` bytes left when parsing the OID.
When we read from the list which `limit_list()` gives us, we need to check that
the commit is still interesting, as it might have become uninteresting after it
was added to the list.
`git-rebase--merge` does not ask for time sorting, but uses the default. We now
produce the same default time-ordered output as git, so make us of that since
it's not always the same output as our time sorting.
It changed from implementation-defined to git's default sorting, as there are
systems (e.g. rebase) which depend on this order. Also specify more explicitly
how you can get git's "date-order".
After `limit_list()` we already have the list in time-sorted order, which is
what we want in the "default" case. Enqueueing into the "unsorted" list would
just reverse it, and the topological sort will do its own sorting if it needs
to.
We've now moved to code that's closer to git and produces the output
during the preparation phase, so we no longer process the commits as
part of generating the output.
This makes a chunk of code redundant, as we're simply short-circuiting
it by detecting we've processed the commits alrady.
After porting over the commit hiding and selection we were still left
with mistmaching output due to the topologial sort.
This ports the topological sorting code to make us match with our
equivalent of `--date-order` and `--topo-order` against the output
from `rev-list`.
This is a convenience function to reverse the contents of a vector and a pqueue
in-place.
The pqueue function is useful in the case where we're treating it as a
LIFO queue.
We had some home-grown logic to figure out which objects to show during
the revision walk, but it was rather inefficient, looking over the same
list multiple times to figure out when we had run out of interesting
commits. We now use the lists in a smarter way.
We also introduce the slop mechanism to determine when to stpo
looking. When we run out of interesting objects, we continue preparing
the walk for another 5 rounds in order to make it less likely that we
miss objects in situations with complex graphs.
When trying to determine if we can safely overwrite an existing workdir
item, we may need to calculate the oid for the workdir item to determine
if its identical to the old side (and eligible for removal).
We previously did this regardless of the type of entry in the workdir;
if it was a directory, we would open(2) it and then try to read(2).
The read(2) of a directory fails on many platforms, so we would treat it
as if it were unmodified and continue to perform the checkout.
On FreeBSD, you _can_ read(2) a directory, so this pattern failed. We
would calculate an oid from the data read and determine that the
directory was modified and would therefore generate a checkout conflict.
This reliance on read(2) is silly (and was most likely accidentally
giving us the behavior we wanted), we should be explicit about the
directory test.
When creating and printing diffs, deal with binary deltas that have
binary data specially, versus diffs that have a binary file but lack the
actual binary data.
Instead of skipping printing a binary diff when there is no data, skip
printing when we have a status of `UNMODIFIED`. This is more in-line
with our internal data model and allows us to expand the notion of
binary data.
In the future, there may have no data because the files were unmodified
(there was no data to produce) or it may have no data because there was
no data given to us in a patch. We want to treat these cases
separately.
When generating diffs for binary files, we load and decompress
the blobs in order to generate the actual diff, which can be very
costly. While we cannot avoid this for the case when we are
called with the `GIT_DIFF_SHOW_BINARY` flag, we do not have to
load the blobs in the case where this flag is not set, as the
caller is expected to have no interest in the actual content of
binary files.
Fix the issue by only generating a binary diff when the caller is
actually interested in the diff. As libgit2 uses heuristics to
determine that a blob contains binary data by inspecting its size
without loading from the ODB, this saves us quite some time when
diffing in a repository with binary files.
According to the reference the git_checkout_tree and git_checkout_head
functions should accept NULL in the opts field
This was broken since the opts field was dereferenced and thus lead to a
crash.
Introduce GIT_OPT_ENABLE_SYMBOLIC_REF_TARGET_VALIDATION option.
Setting this option to 0 allows
validation of a symbolic ref's target to be bypassed.
This option is enabled by default.
This mechanism is added primarily to address a discrepancy between git
behaviour and libgit2 behaviour, whereby the former allows the symbolic
ref target to carry an arbitrary string and the latter does not, so:
$ git symbolic-ref refs/heads/foo bar
$ cat .git/refs/heads/foo
ref: bar
where as attempting the same via libgit2 raises an error:
The given reference name 'bar' is not valid
this mechanism also allows those that might want to make use of
git's more lenient treatment of symbolic ref targets to do so.
When calling `http_connect` on a subtransport whose stream is already
connected, we first close the stream in case no keep-alive is in use.
When doing so, we do not reset the transport's connection state,
though. Usually, this will do no harm in case the subsequent connect
will succeed. But when the connection fails we are left with a
substransport which is tagged as connected but which has no valid
stream attached.
Fix the issue by resetting the subtransport's connected-state when
closing its stream in `http_connect`.
The .gitignore file allows for patterns which unignore previous
ignore patterns. When unignoring a previous pattern, there are
basically three cases how this is matched when no globbing is
used:
1. when a previous file has been ignored, it can be unignored by
using its exact name, e.g.
foo/bar
!foo/bar
2. when a file in a subdirectory has been ignored, it can be
unignored by using its basename, e.g.
foo/bar
!bar
3. when all files with a basename are ignored, a specific file
can be unignored again by specifying its path in a
subdirectory, e.g.
bar
!foo/bar
The first problem in libgit2 is that we did not correctly treat
the second case. While we verified that the negative pattern
matches the tail of the positive one, we did not verify if it
only matches the basename of the positive pattern. So e.g. we
would have also negated a pattern like
foo/fruz_bar
!bar
Furthermore, we did not check for the third case, where a
basename is being unignored in a certain subdirectory again.
Both issues are fixed with this commit.
Support reading and writing index v4. Index v4 uses a very simple
compression scheme for pathnames, but is otherwise similar to index v3.
Signed-off-by: David Turner <dturner@twitter.com>
When failing to initialize a new stransport stream, we try to
release already allocated memory by calling out to
`git_stream_free`, which in turn called out to the stream's
`free` function pointer. As we only initialize the function
pointer later on, this leads to a `NULL` pointer exception.
Furthermore, plug another memory leak when failing to create the
SSL context.
Only provide the empty tree internally, which matches git's behavior.
If we provide the empty blob then any users trying to write it with
libgit2 would omit it from actually landing in the odb, which appear
to git proper as a broken repository (missing that object).
The `SSLCopyPeerTrust` call can succeed but fail to return a trust
object if it can't load the certificate chain and thus cannot check the
validity of a certificate. This can lead to us calling `CFRelease` on a
`NULL` trust object, causing a crash.
Handle this by returning ECERTIFICATE.
Since writing multiple objects may all already exist in a single
packfile, avoid freshening that packfile repeatedly in a tight loop.
Instead, only freshen pack files every 2 seconds.
Don't try to determine when sysdirs are uninitialized. Instead, simply
initialize them all at `git_libgit2_init` time and never try to
reinitialize, except when consumers explicitly call `git_sysdir_set`.
Looking at the buffer length is especially problematic, since there may
no appropriate path for that value. (For example, the Windows-specific
programdata directory has no value on non-Windows machines.)
Previously we would continually trying to re-lookup these values,
which could get racy if two different threads are each calling
`git_sysdir_get` and trying to lookup / clear the value simultaneously.
According to git-fetch(1), "[t]he colon can be omitted when <dst>
is empty." So according to git, the refspec "refs/heads/master"
is the same as the refspec "refs/heads/master:" when fetching
changes. When trying to fetch from a remote with a trailing
colon with libgit2, though, the fetch actually fails while it
works when the trailing colon is left out. So obviously, libgit2
does _not_ treat these two refspec formats the same for fetches.
The problem results from parsing refspecs, where the resulting
refspec has its destination set to an empty string in the case of
a trailing colon and to a `NULL` pointer in the case of no
trailing colon. When passing this to our DWIM machinery, the
empty string gets translated to "refs/heads/", which is simply
wrong.
Fix the problem by having the parsing machinery treat both cases
the same for fetch refspecs.
After 1cd65991, we were passing a pointer to an `unsigned long` to
a function that now expected a pointer to a `size_t`. These types
differ on 64-bit Windows, which means that we trash the stack.
Use `size_t`s in the packbuilder to avoid this.
Somehow I ended up with the following in my ~/.gitconfig:
[branch "master"]
remote = origin
merge = master
rebase = true
I assume something went crazy while I was running the git.git tests
some time ago, and that I never noticed until now.
This is not a good configuration, but it shouldn't cause problems. But
it does. Specifically, if you have this in your config, and you
perform the following set of actions:
create a remote
fetch from that remote
create a branch off of the remote master branch called "master"
delete the branch
delete the remote
The remote delete fails with the message "Could not find key
'branch.master.rebase' to delete". This is because it's iterating over
the config entries (including the ones in the global config) and
believes that there is a master branch which must therefore have these
config keys.
https://github.com/libgit2/libgit2/issues/3856
Ensure that we include conflicts when calling `git_index_read_index`,
which will remove conflicts in the index that do not exist in the new
target, and will add conflicts from the new target.
Most of `git_index_read_index` is common to reading any iterator.
Refactor it out in case we want to implement `read_tree` in terms of it
in the future.
When we create a blame origin, we try to look up the blob that is
to be blamed at a certain revision. When this lookup fails, e.g.
because the file did not exist at that certain revision, we fail
to create the blame origin and return `NULL`. The blame origin
that we have just allocated is thereby free'd with
`origin_decref`.
The `origin_decref` function does not only decrement reference
counts for the blame origin, though, but also for its commit and
blob. When this is done in the error case, we will cause an
uneven reference count for these objects. This may result in
hard-to-debug failures at seemingly unrelated code paths, where
we try to access these objects when they in fact have already
been free'd.
Fix the issue by refactoring `make_origin` such that we only
allocate the object after the only function that may fail so that
we do not have to call `origin_decref` at all. Also fix the
`pass_blame` function, which indirectly calls `make_origin`, to
free the commit when `make_origin` failed.
When showing copy information because we are duplicating contents,
for example, when performing a `diff --find-copies-harder -M100 -B100`,
then show copy from/to lines in a patch, and do not show context.
Ensure that we can also parse such patches.
find_repo had a complex loop and heavily nested conditionals, making it
difficult to follow. Simplify this as much as possible:
- Separate assignments from conditionals.
- Check the complex loop condition in the only place it can change.
- Break out of the loop on error, rather than going through the rest of
the loop body first.
- Handle error cases by immediately breaking, rather than nesting
conditionals.
- Free repo_link unconditionally on the way out of the function, rather
than in multiple places.
- Add more comments on the remaining complex steps.
git_repository_open_ext provides parameters for the start path, whether
to search across filesystems, and what ceiling directories to stop at.
git commands have standard environment variables and defaults for each
of those, as well as various other parameters of the repository. To
avoid duplicate environment variable handling in users of libgit2, add a
GIT_REPOSITORY_OPEN_FROM_ENV flag, which makes git_repository_open_ext
automatically handle the appropriate environment variables. Commands
that intend to act just like those built into git itself can use this
flag to get the expected default behavior.
git_repository_open_ext with the GIT_REPOSITORY_OPEN_FROM_ENV flag
respects $GIT_DIR, $GIT_DISCOVERY_ACROSS_FILESYSTEM,
$GIT_CEILING_DIRECTORIES, $GIT_INDEX_FILE, $GIT_NAMESPACE,
$GIT_OBJECT_DIRECTORY, and $GIT_ALTERNATE_OBJECT_DIRECTORIES. In the
future, when libgit2 gets worktree support, git_repository_open_env will
also respect $GIT_WORK_TREE and $GIT_COMMON_DIR; until then,
git_repository_open_ext with this flag will error out if either
$GIT_WORK_TREE or $GIT_COMMON_DIR is set.
GIT_REPOSITORY_OPEN_NO_SEARCH does not search up through parent
directories, but still tries the specified path both directly and with
/.git appended. GIT_REPOSITORY_OPEN_BARE avoids appending /.git, but
opens the repository in bare mode even if it has a working directory.
To support the semantics git uses when given $GIT_DIR in the
environment, provide a new GIT_REPOSITORY_OPEN_NO_DOTGIT flag to not try
appending /.git.
git only checks ceiling directories when its search ascends to a parent
directory. A ceiling directory matching the starting directory will not
prevent git from finding a repository in the starting directory or a
parent directory. libgit2 handled the former case correctly, but
differed from git in the latter case: given a ceiling directory matching
the starting directory, but no repository at the starting directory,
libgit2 would stop the search at that point rather than finding a
repository in a parent directory.
Test case using git command-line tools:
/tmp$ git init x
Initialized empty Git repository in /tmp/x/.git/
/tmp$ cd x/
/tmp/x$ mkdir subdir
/tmp/x$ cd subdir/
/tmp/x/subdir$ GIT_CEILING_DIRECTORIES=/tmp/x git rev-parse --git-dir
fatal: Not a git repository (or any of the parent directories): .git
/tmp/x/subdir$ GIT_CEILING_DIRECTORIES=/tmp/x/subdir git rev-parse --git-dir
/tmp/x/.git
Fix the testsuite to test this case (in one case fixing a test that
depended on the current behavior), and then fix find_repo to handle this
case correctly.
In the process, simplify and document the logic in find_repo():
- Separate the concepts of "currently checking a .git directory" and
"number of iterations left before going further counts as a search"
into two separate variables, in_dot_git and min_iterations.
- Move the logic to handle in_dot_git and append /.git to the top of the
loop.
- Only search ceiling_dirs and find ceiling_offset after running out of
min_iterations; since ceiling_offset only tracks the longest matching
ceiling directory, if ceiling_dirs contained both the current
directory and a parent directory, this change makes find_repo stop the
search at the parent directory.
Avoid declaring old-style functions without any parameters.
Functions not accepting any parameters should be declared with
`void fn(void)`. See ISO C89 $3.5.4.3.
The old pthread-file did re-implement the pthreads API with exact symbol
matching. As the thread-abstraction has now been split up between Unix- and
Windows-specific files within the `git_` namespace to avoid symbol-clashes
between libgit2 and pthreads, the rewritten wrappers have nothing to do with
pthreads anymore.
Rename the Windows-specific pthread-files to honor this change.
