Mach-O Mach-O Mangler

This is a little library for mangling Mach-O and PE files in various ways. These are the formats used for executables and shared libraries on MacOS and Windows, respectively. (If you want the equivalent for for Linux, then check out patchelf.)

Macho-O features

Some rather specialized (and complex) Mach-O mangling tools designed to support the pynativelib proposal to allow native libraries to be distributed as standalone wheel files. Specifically this includes:

• For pynativelib libraries: a tool that takes a dylib, and a mangling rule, and applies the mangling rule to all the exported symbols. E.g., it can convert a library that exports SSL_new into one that exports pynativelib_openssl__SSL_new. It also changes the library id while it's at it, e.g. from ssl.dylib -> pynativelib_openssl__ssl.dylib (like install_name_tool -id)

  Additionally: a tool that creates a "placeholder" library, which imports the mangled library described above, and then re-exports the symbols under their original names.

• For code that wants to use a pynativelib library: a tool that takes a dylib/bundle/executable, a list of "original" dylibs, and for each "original" dylib, a newname for that dylib, and a mangling rule. It then (a) replaces the import of the original dylib with an absolute import of the new dylib name from a non-existent directory, (b) marks this as a "weak" import, (c) applies the mangling rule to all symbols imported from this dylib, (d) marks these symbols for lookup in the flat namespace.

It turns out that this exact combination of things is the only way provided for by the MacOS linker/loader to have dylib/bundle A linked against dylib B where the relative on-disk location of A and B is not known until after the executable starts, while preserving the usual two-level namespace rules for avoiding symbol collisions. I promise it will all make sense once I have a chance to write it up properly...

Some known limitations of the Mach-O mangling code:

• Unsurprisingly, this kind of patching does not play well with code signing. The code doesn't take any special case with signatures; they'll probably just get messed up. If you want to sign your binaries, then do your mangling first before signing.
• We currently only rewrite the new-style DYLD_INFO symbol table (introduced in 10.5), not the (almost?) totally redundant SYMTAB/DYSYMTAB symbol table. (Interesting fact: all Mach-O binaries include two completely different representations of their symbols tables. The new one is more compact, to save space, but then they keep the old one around for compatibility, so... anyway.) As far as I can tell, the only thing in in modern MacOS that still uses SYMTAB/DYSYMTAB is dladdr, and I don't think anyone is relying on dladdr output for, well... anything? I think worst case, you might end up seeing the original symbol names inside a debugger or profiler? But this wouldn't be too hard to fix if it becomes a problem.
• It doesn't do any special handling of the DYLD_INFO weak_bind table, or weak exports. (NB these have nothing to do __attribute__((weak)) or __attribute__((weak_import)) or any of the mentions of the word weak in the ld man page – I think they're for implementing vague linkage.) This is probably not a disastrous option, but I'm not 100% sure whether it's actually correct – it's an incredibly obscure part of the Mach-O format, and Mach-O is pretty obscure to start with. Fortunately this feature is only used by C++ libraries, so we can get started without it.
• When mangling imports, we convert any lazy imports (that need mangling) into eager imports. This is required because the lazy import stubs hard-code the memory layout of the import table into immediate constants inside the stub assembly itself, and I do not feel like trying to automatically rewrite x86-64 opcodes. Instead, we leave the lazy import table alone (so all the unmangled lazy imports can continue to use it), and eagerly bind all the mangled imports, so the unmangled stubs never get called.
• I noticed some new code dyld in MacOS 10.12 that imposes some annoying arbitrary restrictions on which order the different bits of DYLD_INFO appear in the file. This should only affect libraries that are built with 10.12 as their minimum required version, so for folks trying to build stuff for general distribution this shouldn't matter for a while. This also isn't hard to fix, it just means that we'll probably have to start making some pointless redundant copies of bits of the file that we didn't change, just so that the second copy can be placed after the bit of the file that we did change, which is tiresome and I haven't gotten around to it yet.
• When mangling imports, we don't check for re-exports, which are also a kind of import. Should probably fix this...

PE features

A tool that can read in a PE file (.exe or .dll) that is currently linked to foo.dll, and rewrite it so that it becomes linked to bar.dll instead (similar to patchelf --replace on Linux, or install_name_tool -change on OS X). This is useful for avoiding naming collisions between different versions of the same library.

For example, suppose you have two Python extensions A.dll and B.dll, that are distributed separately by different people. They both contain some fortran code linked to to libgfortran-3.dll, so both packages ship a copy of libgfortran-3.dll. Because of the way Windows DLL loading works, what will happen is that if I load A.dll first, then both A.dll and B.dll will end up using A's copy of libgfortran-3.dll, while B's copy will be ignored. (Or vice-versa if I import B first.) This will happen even if I arrange things so that A's copy is not on the DLL search path at the time that B is loaded -- Windows always checks for already-loaded DLL's with a given basename before it actually checks the DLL search path (modulo some complications around SxS assemblies, but you don't really want to go there).

This is bad, because there's no guarantee that B.dll will work with A's version of libgfortran-3.dll (e.g., A's copy might be too old for B). Welcome to DLL hell!

We could avoid all this by renaming the colliding libraries to have different names, e.g. libgfortran-3-for-A.dll and libgfortran-3-for-B.dll. But if we just rename the files, then everything will break, because A.dll is looking for libgfortran-3.dll, not libgfortran-3-for-A.dll.

This is where machomachomangler comes in: it lets you patch A.dll so that it's linked to libgfortran-3-for-A.dll. And then everything works. Hooray.

This basically solves the same problem as private SxS assemblies, except better in all ways: it's simpler (no XML manifests), more flexible (no finicky requirements for the filesystem layout), and doesn't require reading the awful SxS assembly documentation.

Example usage:

$ python3 -m machomachomangler.cmd.redll A.dll A-patched.dll libgfortran-3.dll libgfortan-3-for-A.dll

There's an example in example/ then you can play with. E.g. on Debian with a mingw-w64 cross-compiler and wine installed:

$ cd pe-example/

$ ./build.sh
+ i686-w64-mingw32-gcc -shared test_dll.c -o test_dll.dll
+ i686-w64-mingw32-gcc test.c -o test.exe -L. -ltest_dll
+ i686-w64-mingw32-strip test.exe

$ wine test.exe
dll_function says: test_dll

$ mv test_dll.dll test_dll_renamed.dll

# Apparently wine's way of signalling a missing DLL is to fail silently.
$ wine test.exe || echo "failed -- test_dll.dll is missing"
failed -- test_dll.dll is missing

$ PYTHONPATH=.. python3 -m machomachomangler.cmd.redll test.exe test-patched.exe test_dll.dll test_dll_renamed.dll

# Now it works again:
$ wine test-patched.exe
dll_function says: test_dll

Some known limitations of the PE dll-import-switcheroo code:

• The command line tool could be less minimalist.

• GNU objdump has a bug where it can't read the import tables of our patched PE files -- it just shows all of the import table until it hits the patched entry, and then it stops displaying anything. (The issue is that binutils wants all the data involved in the import tables to come from a single PE section.) However, I've tried giving the patched files to Dependency Walker, wine, and Windows itself, and they all handle them fine -- so the files are okay, it's just a bug in objdump. Just be warned that if you're trying to use objdump to check if the patching worked, then it's almost certainly going to tell you a confusing lie.

• Unsurprisingly, this kind of patching does not play well with code signing. We try to at least clear any existing signatures (so that the binary becomes unsigned, rather than signed with an invalid signature), but this hasn't been tested.

• We don't try to handle files with trailing data after the end of the PE file proper. This commonly occurs with e.g. self-extracting archives and installers. Shouldn't be a big deal in theory, but I did find that when compiling a simple .exe with mingw-w64 the tool refused to work until I had run strip on the binary, even though in theory this should work fine -- so probably there's some improvements possible.

  [Note to self: it looks like this is a GNU extension for putting long section names into PE files, which I guess are they use for their debug format -- this is documented here, search for "Coff long section names". It's probably not hard to handle this better, e.g. by stripping it ourself or even fixing it up.]

• We don't try to update the PE header checksum, since the algorithm for doing this is (nominally) a secret, and I'm informed that for regular user-space code there's nothing that actually cares about whether it's correct. But my information could be wrong. (Note: it looks like binutils might know how to compute this checksum? I'm not sure.)

  [Update: Stefan Kanthak informs me that this algorithm is well known, and in fact it looks pefile has an MIT-licensed Python implementation so I guess it might be good to fix this at some point.]

General limitations

Only tested on Python 3.4 and 3.5. Probably any Python 3 will work, and Python 2 definitely won't without some fixes. (There's lots of fiddly byte-string handling.)

I'm lazy, so I just load the whole binary files into memory -- maybe several copies of it. This actually wouldn't be too hard to fix (using memory mapping etc.) but I guess it doesn't matter that much because who has multi-gigabyte Mach-O/PE images??

Contact

wheel-builders@python.org

License

It's Saturday afternoon, I've got the flu or something, and I'm spending my free time writing software to make some proprietary operating systems -- ones that are backed by one of the world's larger corporations -- better able to compete for developers with other, better-designed operating systems. I mean, I'm not saying that poring over the PE/COFF specification isn't fun! But it's not that fun. (And honestly the Mach-O docs are absolutely terrible, to the extent they exist at all.)

To assuage my annoyance, this software is licensed under the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License or (at your option) any later version. See LICENSE.txt for details.

This shouldn't have any effect on most uses, since it only affects people who are redistributing this software or running it on behalf of other people; you can use this software to manipulate your BSD-licensed DLLs, your proprietary-licensed DLLs, or whatever you like, and that's fine. The license affects the code for machomachomangler itself; not the code you run it on.

However, if for some reason you or your company have some kind of allergy to this license, send me an email and we'll work out an appropriate tithe.

Also, to preserve our options in case I get over this fit of pique, please license all contributions under the MIT license. (I definitely will not switch to any proprietary license, but might switch to a permissive OSS license.) Thanks!

Code of conduct

Contributors are requested to follow our code of conduct in all project spaces.