docs: explicitly document the threading caveatscmn/threading

Talk about sharing objects and error messages; but the most important
part is about what to do with the cryptographic libraries, which sadly
have to become to responsibility of the application.

2 files changed, 84 insertions, 0 deletions

diff --git a/README.md b/README.md
index 98aaa1dd7..f25bb12bb 100644
--- a/README.md
+++ b/README.md
@@ -58,6 +58,11 @@ dependencies, it can make use of a few libraries to add to it:
 - LibSSH2 to enable the SSH transport
 - iconv (OSX) to handle the HFS+ path encoding peculiarities
 
+Threading
+=========
+
+See [THREADING](THREADING.md) for information
+
 Building libgit2 - Using CMake
 ==============================
 
diff --git a/THREADING.md b/THREADING.md
new file mode 100644
index 000000000..475931885
--- /dev/null
+++ b/THREADING.md
@@ -0,0 +1,79 @@
+Threads in libgit2
+==================
+
+You may safely use any libgit2 object from any thread, though there
+may be issues depending on the cryptographic libraries libgit2 or its
+dependencies link to (more on this later). For libgit2 itself,
+provided you take the following into consideration you won't run into
+issues:
+
+Sharing objects
+---------------
+
+Use an object from a single thread at a time. Most data structures do
+not guard against concurrent access themselves. This is because they
+are rarely used in isolation and it makes more sense to synchronize
+access via a larger lock or similar mechanism.
+
+There are some objects which are read-only/immutable and are thus safe
+to share across threads, such as references and configuration
+snapshots.
+
+Error messages
+--------------
+
+The error message is thread-local. The `giterr_last()` call must
+happen on the same thread as the error in order to get the
+message. Often this will be the case regardless, but if you use
+something like the GDC on MacOS (where code is executed on an
+arbitrary thread), the code must make sure to retrieve the error code
+on the thread where the error happened.
+
+Threads and cryptographic libraries
+=======================================
+
+On Windows
+----------
+
+When built as a native Windows DLL, libgit2 uses WinCNG and WinHTTP,
+both of which are thread-safe. You do not need to do anything special.
+
+When using libssh2 which itself uses WinCNG, there are no special
+steps necessary. If you are using a MinGW or similar environment where
+libssh2 uses OpenSSL or libgcrypt, then the non-Windows case affects
+you.
+
+Non-Windows
+-----------
+
+On the rest of the platforms, libgit2 uses OpenSSL to be able to use
+HTTPS as a transport. This library is made to be thread-implementation
+agnostic, and the users of the library must set which locking function
+it should use. This means that libgit2 cannot know what to set as the
+user of libgit2 may use OpenSSL independently and the locking settings
+must survive libgit2 shutting down.
+
+libgit2 does provide a convenience function
+`git_openssl_set_locking()` to use the platform-native mutex
+mechanisms to perform the locking, which you may rely on if you do not
+want to use OpenSSL outside of libgit2, or you know that libgit2 will
+outlive the rest of the operations. It is not safe to use OpenSSL
+multi-threaded after libgit2's shutdown function has been called.
+
+See the
+[OpenSSL documentation](https://www.openssl.org/docs/crypto/threads.html)
+on threading for more details.
+
+libssh2 may be linked against OpenSSL or libgcrypt. If it uses
+OpenSSL, you only need to set up threading for OpenSSL once and the
+above paragraphs are enough. If it uses libgcrypt, then you need to
+set up its locking before using it multi-threaded. libgit2 has no
+direct connection to libgcrypt and thus has not convenience functions for
+it (but libgcrypt has macros). Read libgcrypt's
+[threading documentation for more information](http://www.gnupg.org/documentation/manuals/gcrypt/Multi_002dThreading.html)
+
+It is your responsibility as an application author or packager to know
+what your dependencies are linked against and to take the appropriate
+steps to ensure the cryptographic libraries are thread-safe. We agree
+that this situation is far from ideal but at this time it is something
+the application authors need to deal with.