Previous to pulling out and extending the fake backend, it was quite
cumbersome to write tests for very specific scenarios regarding
backends. But as we have made it more generic, it has become much easier
to do so. As such, this commit adds multiple tests for scenarios with
multiple backends for the ODB.

The changes also include a test for a very targeted scenario. When one
backend found a matching object via `read_prefix`, but the last backend
returns `GIT_ENOTFOUND` and when object hash verification is turned off,
we fail to reset the error code to `GIT_OK`. This causes us to segfault
later on, when doing a double-free on the returned object.

Right now, the fake backend is quite restrained in the way how it
works: we pass it an OID which it is to return later as well as an error
code we want it to return. While this is sufficient for existing tests,
we can make the fake backend a little bit more generic in order to allow
us testing for additional scenarios.

To do so, we change the backend to not accept an error code and OID
which it is to return for queries, but instead a simple array of OIDs
with their respective blob contents. On each query, the fake backend
simply iterates through this array and returns the first matching
object.

In order to make the fake backend more useful, we want to enable it
holding multiple object references. To do so, we need to decouple it
from the single fake OID it currently holds, which we simply move up
into the calling tests.

The fake backend used by the test suite `odb::backend::nonrefreshing` is
useful to have some low-level tests for the ODB layer. As such, we move
the implementation into its own `backend_helpers` module.

smart_protocol: fix parsing of server ACK responses

odb_read_prefix: reset error in backends loop

When looking for an object by prefix, we query all the backends so that
we can ensure that there is no ambiguity.  We need to reset the `error`
value between backends; otherwise the first backend may find an object
by prefix, but subsequent backends may not.  If we do not reset the
`error` value then it will remain at `GIT_ENOTFOUND` and `read_prefix_1`
will fail, despite having actually found an object.

Update version number to 0.26

CHANGELOG: add various changes introduced since v0.25

Ensure packfiles with different contents have different names

Update to forced checkout and untracked files

Only ignore `EBUSY` from `rmdir` when the `GIT_RMDIR_SKIP_NONEMPTY` bit
is set.

When deleting a directory during checkout, do not simply delete the
directory, since there may be untracked files.  Instead, go into
the iterator and examine each file.

In the original code (the code with the faulty assumption), we look to
see if there's an index entry beneath the directory that we want to
remove.   Eg, it looks to see if we have a workdir entry foo and an
index entry foo/bar.txt. If this is not the case, then the working
directory must have precious files in that directory. This part is okay.
The part that's not okay is if there is an index entry foo/bar.txt. It
just blows away the whole damned directory.

That's not cool.

Instead, by simply pushing the directory itself onto the stack and
iterating each entry, we will deal with the files one by one - whether
they're in the index (and can be force removed) or not (and are
precious).

The original code was a bad optimization, assuming that we didn't need
to git_iterator_advance_into if there was any index entry in the folder.
That's wrong - we could have optimized this iff all folder entries are
in the index.

Instead, we need to simply dig into the directory and analyze its
entries.

If the `GIT_CHECKOUT_FORCE` flag is given to any of the `git_checkout`
invocations, we remove files which were previously staged. But while
doing so, we unfortunately also remove unstaged files in a directory
which contains at least one staged file, resulting in potential data
loss.

This commit adds two tests to verify behavior.

Fix ACK parsing in wait_while_ack() internal function. This patch
handles the case where multi_ack_detailed mode sends 'ready' ACKs. The
existing functionality would bail out too early, thus causing the
processing of the ensuing packfile to fail if/when 'ready' ACKs were
sent.

settings: rename `GIT_OPT_ENABLE_SYNCHRONOUS_OBJECT_CREATION`

Initially, the setting has been solely used to enable the use of
`fsync()` when creating objects. Since then, the use has been extended
to also cover references and index files. As the option is not yet part
of any release, we can still correct this by renaming the option to
something more sensible, indicating not only correlation to objects.

This commit renames the option to `GIT_OPT_ENABLE_FSYNC_GITDIR`. We also
move the variable from the object to repository source code.

Buffer growing cleanups

Coverity fixes

The `git_buf_init` function has an optional length parameter, which will
cause the buffer to be initialized and allocated in one step. This can
be used instead of static initialization with `GIT_BUF_INIT` followed by
a `git_buf_grow`. This patch does so for two functions where it is
applicable.

Both the `git_buf_init` and `git_buf_attach` functions may call
`git_buf_grow` in case they were given an allocation length as
parameter. As such, it is possible for these functions to fail when we
run out of memory. While it won't probably be used anytime soon, it does
indeed make sense to also record this fact by returning an error code
from both functions. As they belong to the internal API only, this
change does not break our interface.

The `ENSURE_SIZE` macro can be used to grow a buffer if its currently
allocated size does not suffice a required target size. While most of
the code already uses this macro, the `git_buf_join` and `git_buf_join3`
functions do not yet use it. Due to the macro first checking whether we
have to grow the buffer at all, this has the benefit of saving a
function call when it is not needed. While this is nice to have, it will
probably not matter at all performance-wise -- instead, this only serves
for consistency across the code.

While the `ENSURE_SIZE` macro gets a reference to both the buffer that
is to be resized and a new size, we were not consistently referencing
the passed buffer, but instead a variable `buf`, which is not passed in.
Funnily enough, we never noticed because our buffers seem to always be
named `buf` whenever the macro was being used.

Fix the macro by always using the passed-in buffer. While at it, add
braces around all mentions of passed-in variables as should be done with
macros to avoid subtle errors.

Found-by: Edward Thompson

The function `git_buf_try_grow` consistently calls `giterr_set_oom`
whenever growing the buffer fails due to insufficient memory being
available. So in fact, we do not have to do this ourselves when a call
to any buffer-growing function has failed due to an OOM situation. But
we still do so in two functions, which this patch cleans up.

SHA1DC update

The updated SHA1DC library allows us to use custom includes instead of
using standard includes. Due to requirements with cross-platform, we
provide some custom system includes files like for example the
"stdint.h" file on Win32. Because of this, we want to make sure to avoid
breaking cross-platform compatibility when SHA1DC is enabled.

To use the new mechanism, we can simply define
`SHA1DC_NO_STANDARD_INCLUDES`. Furthermore, we can specify custom
include files via two defines, which we now use to include our
"common.h" header.

This updates our version of SHA1DC to e139984 (Merge pull request #35
from lidl/master, 2017-05-30).

Fix path computations for compressed index entries

(Temporarily) disable UNC tests

(Temporarily) disable UNC path tests to work around AppVeyor issues.

Fix build with libressl

OpenSSL v1.1 has introduced a new way of initializing the library
without having to call various functions of different subsystems. In
libgit2, we have been adapting to that change with 88520151
(openssl_stream: use new initialization function on OpenSSL version
>=1.1, 2017-04-07), where we added an #ifdef depending on the OpenSSL
version. This change broke building with libressl, though, which has not
changed its API in the same way.

Fix the issue by expanding the #ifdef condition to use the old way of
initializing with libressl.

Signed-off-by: Marc-Antoine Perennou <Marc-Antoine@Perennou.com>