When computing renames, we cache the hash signatures for each of the
potentially conflicting entries so that we do not need to repeatedly
read the file and can at least halfway efficiently determine whether two
files are similar enough to be deemed a rename. In order to make the
hash signatures meaningful, we require at least four lines of data to be
present, resulting in at least four different hashes that can be
compared. Files that are deemed too small are not cached at all and
will thus be repeatedly re-hashed, which is usually not a huge issue.

The issue with above heuristic is in case a file does _not_ have at
least four lines, where a line is anything separated by a consecutive
run of "\n" or "\0" characters. For example "a\nb" is two lines, but
"a\0\0b" is also just two lines. Taken to the extreme, a file that has
megabytes of consecutive space- or NUL-only may also be deemed as too
small and thus not get cached. As a result, we will repeatedly load its
blob, calculate its hash signature just to finally throw it away as we
notice it's not of any value. When you've got a comparitively big file
that you compare against a big set of potentially renamed files, then
the cost simply expodes.

The issue can be trivially fixed by introducing negative cache entries.
Whenever we determine that a given blob does not have a meaningful
representation via a hash signature, we store this negative cache marker
and will from then on not hash it again, but also ignore it as a
potential rename target. This should help the "normal" case already
where you have a lot of small files as rename candidates, but in the
above scenario it's savings are extraordinarily high.

To verify we do not hit the issue anymore with described solution, this
commit adds a test that uses the exact same setup described above with
one 50 megabyte blob of '\0' characters and 1000 other files that get
renamed. Without the negative cache:

$ time ./libgit2_clar -smerge::trees::renames::cache_recomputation >/dev/null
real    11m48.377s
user    11m11.576s
sys     0m35.187s

And with the negative cache:

$ time ./libgit2_clar -smerge::trees::renames::cache_recomputation >/dev/null
real    0m1.972s
user    0m1.851s
sys     0m0.118s

So this represents a ~350-fold performance improvement, but it obviously
depends on how many files you have and how big the blob is. The test
number were chosen in a way that one will immediately notice as soon as
the bug resurfaces.

refdb_backend: improve callback documentation

credentials: provide backcompat for opaque structs

The credential structures are now opaque and defined in
`sys/credential.h`.  However, we should continue to provide them for
backward compatibility, unless `GIT_DEPRECATED_HARD` is set.

Fix segfault when calling git_blame_buffer()

The callbacks are currently sparsely documented, making it really hard
to implement a new backend without taking a look at the existing
refdb_fs backend. Add documentation to make this task hopefully easier
to achieve.

Fix spelling error

Signed-off-by: Utkarsh Gupta <utkarsh@debian.org>

This change makes sure that the hunk is not null before trying to
dereference it. This avoids segfaults, especially when blaming against a
modified buffer (i.e. the index).

Fixes: #5443

refdb_fs: initialize backend version

repository: improve commondir docs

While the `git_refdb_backend()` struct has a version, we do not
initialize it correctly when calling `git_refdb_backend_fs()`. Fix this
by adding the call to `git_refdb_init_backend()`.

cmake: use install directories provided via GNUInstallDirs

azure: fix errors due to curl and removal of old VM images

mbedTLS has fixed their certificate.  🎉

We currently hand-code logic to configure where to install our artifacts
via the `LIB_INSTALL_DIR`, `INCLUDE_INSTALL_DIR` and `BIN_INSTALL_DIR`
variables. This is reinventing the wheel, as CMake already provide a way
to do that via `CMAKE_INSTALL_<DIR>` paths, e.g. `CMAKE_INSTALL_LIB`.
This requires users of libgit2 to know about the discrepancy and will
require special hacks for any build systems that handle these variables
in an automated way. One such example is Gentoo Linux, which sets up
these paths in both the cmake and cmake-utils eclass.

So let's stop doing that: the GNUInstallDirs module handles it in a
better way for us, especially so as the actual values are dependent on
CMAKE_INSTALL_PREFIX. This commit removes our own set of variables and
instead refers users to use the standard ones.

As a second benefit, this commit also fixes our pkgconfig generation to
use the GNUInstallDirs module. We had a bug there where we ignored the
CMAKE_INSTALL_PREFIX when configuring the libdir and includedir keys, so
if libdir was set to "lib64", then libdir would be an invalid path. With
GNUInstallDirs, we can now use `CMAKE_INSTALL_FULL_LIBDIR`, which
handles the prefix for us.

We currently hve some problems with our curl downloads when building
Docker images. It's not quite obvious what the problem is and they seem
to occur semi-randomly. To unblock our CI, let's add the "--insecure"
flag to curl to ignore any certificate errors. This is intended as a
temporary solution only.

Azure is phasing out old images on March 23rd 2020, but we're currently
still using them. So let's upgrade images as following:

    - Ubuntu 16.04 -> ubuntu-18.04
    - macOS 10.13 -> macOS-10.15
    - Hosted Windows machines -> vs2017-win2016

Each of them is currently the latest version. As the new Microsoft
Windows machine has upgraded to MSVS2017, we need to also adjust our
CMake generators to "Visual Studio 15 2017". As this CMake generator
doesn't accept the target platform name anymore, we instead need to set
it up via either "-A Win32" or "-A x64".

[1]: https://devblogs.microsoft.com/devops/removing-older-images-in-azure-pipelines-hosted-pools/

Our Docurium builds currently depend on Debian Jessie, which has CMake
v3.0 available. As rugged has bumped its CMake requirements to need at
least v3.5 now, the documentation build is thus failing.

Fix this by converting our Docurium Docker image to be based on Ubuntu
Bionic. We already do base all of our images on Ubuntu, so I don't see
any sense in using Debian here. If this was only to speed up builds, we
should just go all the way and use some minimal container like Alpine
anyway.

Also remove cache busters. As we're rebuilding the image every time,
it's we really don't need them at all.

We currently pass the "--silent" flag to most invocations of curl, but
in fact this does not only suppress the progress meter, but also any
errors. So let's also pass "--show-error", too.

The `CC_MD4()` function has been deprecated in macOS 10.15. Silence this
warning for now until we implement a proper fix.

The Secure Transport interface we're currently using has been deprecated
with macOS 10.15. As we're currently in code freeze, we cannot migrate
to newer interfaces. As such, let's disable deprecation warnings for
our "schannel.c" stream.

win32: don't canonicalize relative paths

Ensure that we don't canonicalize symlink targets.

Don't canonicalize symlink targets; our win32 path canonicalization
routines expect an absolute path.  In particular, using the path
canonicalization routines for symlink targets (introduced in commit
7d55bee6, "win32: fix relative symlinks pointing into dirs",
2020-01-10).

Now, use the utf8 -> utf16 relative path handling functions, so that
paths like "../foo" will be translated to "..\foo".

Add a function that takes a (possibly) relative UTF-8 path and emits a
UTF-16 path with forward slashes translated to backslashes.  If the
given path is, in fact, absolute, it will be translated to absolute path
handling rules.

The path canonicalization functions on win32 are intended to
canonicalize absolute paths; those with prefixes.  In other words,
things start with drive letters (`C:\`), share names (`\\server\share`),
or other prefixes (`\\?\`).

This function removes leading `..` that occur after the prefix but
before the directory/file portion (eg, turning `C:\..\..\..\foo` into
`C:\foo`).  This translation is not appropriate for local paths.

Fixes #5428

CMake booleans

Set proper pkg-config dependency for pcre2

httpclient: use a 16kb read buffer for macOS

Use a 16kb read buffer for compatibility with macOS SecureTransport.

SecureTransport `SSLRead` has the following behavior:

1. It will return _at most_ one TLS packet's worth of data, and
2. It will try to give you as much data as you asked for

This means that if you call `SSLRead` with a buffer size that is smaller
than what _it_ reads (in other words, the maximum size of a TLS packet),
then it will buffer that data for subsequent calls.  However, it will
also attempt to give you as much data as you requested in your SSLRead
call.  This means that it will guarantee a network read in the event
that it has buffered data.

Consider our 8kb buffer and a server sending us 12kb of data on an HTTP
Keep-Alive session.  Our first `SSLRead` will read the TLS packet off
the network.  It will return us the 8kb that we requested and buffer the
remaining 4kb.  Our second `SSLRead` call will see the 4kb that's
buffered and decide that it could give us an additional 4kb.  So it will
do a network read.

But there's nothing left to read; that was the end of the data.  The
HTTP server is waiting for us to provide a new request.  The server will
eventually time out, our `read` system call will return, `SSLRead` can
return back to us and we can make progress.

While technically correct, this is wildly ineffecient.  (Thanks, Tim
Apple!)

Moving us to use an internal buffer that is the maximum size of a TLS
packet (16kb) ensures that `SSLRead` will never buffer and it will
always return everything that it read (albeit decrypted).

Signed-off-by: Igor Raits <i.gnatenko.brain@gmail.com>

ci: provide globalsign certs for bionic

tls.mbed.org has neglected to send their full certificate chain.  Add
their intermediate cert manually.  🙄

deps: ntlmclient: fix htonll on big endian FreeBSD

azure-pipelines: download GlobalSign's certificate manually

tls.mbed.org has neglected to send their full certificate chain.  Add
their intermediate cert manually.  🙄

In commit 3828ea67 (deps: ntlmclient: fix missing htonll symbols on
FreeBSD and SunOS, 2020-02-21), we've fixed compilation on BSDs due to
missing `htonll` wrappers. While we are now using `htobe64` for both
Linux and OpenBSD, we decided to use `bswap64` on FreeBSD. While correct
on little endian systems, where we will swap from little- to big-endian,
we will also do the swap on big endian systems. As a result, we do not
use network byte order on such systems.

Fix the issue by using htobe64, as well.

deps: ntlmclient: fix missing htonll symbols on FreeBSD and SunOS