Applications on Windows that don't link to the MSVCR90.DLL via a
manifest are broken with Python 2.6.1. This includes apps that link
with the C library statically and apps that link with other versions of
it. These applications worked fine with Python 2.6.

To test this, I created a simple application that did nothing but call
Py_Main (compiled with VS2008). When that app links with the C library
statically, attempting to import the socket module gives this error:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Python26\lib\socket.py", line 46, in <module>
    import _socket
ImportError: DLL load failed: The specified module could not be found.

When that app links with the C library dynamically, it works properly.

In particular, this issue breaks mod_python. From the Apache error log:

[Sat Dec 06 00:49:21 2008] [error] make_obcallback: could not import
mod_python.apache.\n
Traceback (most recent call last):
  File "C:\Python26\lib\site-packages\mod_python\apache.py", line 29, in
<module>
    import cgi
  File "C:\Python26\Lib\cgi.py", line 40, in <module>
    import urllib
  File "C:\Python26\Lib\urllib.py", line 26, in <module>
    import socket
  File "C:\Python26\lib\socket.py", line 46, in <module>
    import _socket
ImportError: DLL load failed: A dynamic link library (DLL)
initialization routine failed.

I'm guessing this is a side-effect of the fix for bpo-4120. Since modules
like _socket.pyd don't have manifests that tell Windows where to find
the side-by-side assembly to load, and the application loading
_socket.pyd doesn't have it in its manifest either, Windows has no clue
where to find MSVCR90.dll.

Something I discovered (by accident) is that, if there's an external
manifest in the same folder as the hosting application (ie,
testpy.exe.manifest) that points to the C runtime, it fixes the problem,
even if the host app links statically to the C runtime. I don't recall
off-hand whether these external manifests take precedence over the
internal ones, but in this case it seems to provide enough information
to Windows that it can resolve the dependency.

I've attached the test program I was using. The commands I used to
compile it are:

cl /MT /c /IC:\Python26\include testpy.c
link /LIBPATH:C:\Python26\libs testpy.obj

I've attached a manifest file that references the C runtime; adding this
file to the Apache bin folder seems to workaround this issue.

I would claim that this is not a bug. There is simpler no other way to
deal with this entire SxS mess.

A mess it is :/

I understand the rationale behind bpo-4120, but it seems like it only helps
a narrow set of applications, namely "applications that link dynamically
with the same version of MSVCR90 as Python and that bundle the MSVCR90
DLL and that can't install the VS2008 redist". In 2.6.1 those apps
don't have to install the VS2008 redist, but every other app needs to
either bundle the runtime DLLs (as a private assembly) or use the
manifest workaround I described above, even if the VS2008 redist is
installed on the system.

The 2.6.0 behavior - requiring the VS2008 redist to be installed - is
hardly perfect (to put it mildly), but in my opinion it's more obvious
and straightforward, and more consistent with the behavior of other
Windows software.

I understand the rationale behind bpo-4120, but it seems like it only helps
a narrow set of applications, namely "applications that link dynamically
with the same version of MSVCR90 as Python and that bundle the MSVCR90
DLL and that can't install the VS2008 redist".

This is not a narrow set, though. It includes all the applications that
use py2exe and friends to create stand-alone applications. This case
absolutely must be supported.

The 2.6.0 behavior - requiring the VS2008 redist to be installed - is
hardly perfect (to put it mildly), but in my opinion it's more obvious
and straightforward, and more consistent with the behavior of other
Windows software.

Python has a long tradition of supporting "xcopy deployment". I don't
want Microsoft to dictate that we stop supporting that. I find the need
to have end-users install the CRT redist particularly unacceptable.

I don't quite understand this issue yet. python26.dll is linked with
a manifest. Isn't that good enough?

I don't quite understand this issue yet. python26.dll is linked with
a manifest. Isn't that good enough?

Apparently not, at least in my testing. It seems that if a DLL is
loaded, Windows will try to resolve its dependencies by looking at that
DLL's own manifest; if that fails, Windows will try to resolve them by
looking at the EXE's manifest. It won't check other DLLs. From the
loader's perspective, it seems like there's no "tree" of DLLs, it's just
"EXE loads DLL, EXE loads DLL", even if the loading code is actually in
one of the DLLs.

I guess I'm more concerned about applications like Apache that only use
Python through an external module or plugin; there's no reason the
Apache developers would be expected to make this kind of change with the
manifests and everything, much less any commercial, closed-source app
(say, an IDE or editor that has plugins for several scripting
languages). Also, the manifest trick I described as a workaround was
only as simple as it was in this case because httpd.exe didn't have any
manifest at all; if it already had an internal manifest, working-around
this would have been much more gruesome.

I understand, though, the value of "xcopy deployment", and I realize
that having some means of doing that is better than none at all.

Here's an option, though unfortunately not a trivial one: use a private
build of the C runtime. The Windows version of Firefox does this
(mozcrt19.dll). The private CRT build doesn't use SxS in any way, so it
gets around this issue, as well as other issues like not allowing "for
me" installs on Vista (bpo-4018).

For me to build the CRT from the source included with Visual Studio 2008
took some tweaking - apparently having the CRT source build properly out
of the box wasn't a priority for MS (the Mozilla CRT seems to be built
from the VS2005 source; perhaps that version is more cooperative). It
does yield a CRT that links like an ordinary DLL, but the fact that it
can't be built without modifications is a major drawback to this solution.

Also, I'd assume that the CRT source isn't included in the Express
version (and possibly other ones), so that's a downside, too.

As far as Python itself, the project configurations would have to be
changed to define _CRT_NOFORCE_MANIFEST, suppress the default linking to
the stock CRT (/NODEFAULTLIB:msvcrt.lib), and link to the import library
for the private CRT. I might try to create an experimental solution
config for this.

I've no time to dig deeper now as I suspect testing will require removal
of the vc9 assembly from the GAC and testing with a local one, but some
comments:

test.c's error is "can't find the DLL" - this will be as we attempt to
load Python's DLL - but this isn't the same as the original error, which
is "DLL init routine failed". To repro the initial error, I suspect you
will want to put the full assembly next to test.exe - that will allow
python.dll to load - then test.c should call PyExec_EvalString("import
socket\n") - it is at *that* point the error we care about is likely to
be thrown.

That specific error code means the DLL init routine in the CRT started
executing, but explicitly threw an error. Its possible to use the
debugger to see exactly when this is thrown, and it relates to the
assembly configuration (the details escape me). I suspect it will be
necessary to locate this CRT init code to determine exactly why it is
throwing the error, and possibly determine how to satisfy it (the DLL
*is* loaded, so it should be capable of working)

I meant to mention: FWIW, *some* py2exe apps work fine with the old
scheme - specifically, IIUC, any app will work fine so long as the .pyd
files were next to the executable, which is next to the assembly. I
understand this is a significant restriction, but its worth mentioning
for completeness.

test.c's error is "can't find the DLL" - this will be as we attempt to
load Python's DLL - but this isn't the same as the original error, which
is "DLL init routine failed". To repro the initial error, I suspect you
will want to put the full assembly next to test.exe - that will allow
python.dll to load - then test.c should call PyExec_EvalString("import
socket\n") - it is at *that* point the error we care about is likely to
be thrown.

The test program isn't having a problem loading python26.dll - it gets
to an interpreter prompt and it can execute simple Python statements.
It doesn't throw ImportError until the user types in "import socket".
Further, I can see in Process Explorer that python26.dll is loaded in
the running testpy.exe process. I apologize if my initial description
wasn't clear on that point.

I do see the error codes are different (between testpy.c and
mod_python), but both are triggered by trying to load _socket.pyd.

Nonetheless: if I copy the CRT assembly into the same folder as
testpy.exe, I get a popup dialog when I try to import socket, saying
"Runtime Error!" and "R6034; An application has made an attempt to load
the C library incorrectly.". When I click OK in that dialog, I get this
error in the console running testpy.exe:

ImportError: DLL load failed: A dynamic link library (DLL)
initialization routine failed.

(If the assembly is not in the folder, there is no popup dialog at all.)
So it does indeed change the error code received.

Also, I see now that I made a mistake in reporting the error code from
the Apache log. The actual behavior is this:

• Ordinary, the ImportError is:
  ImportError: DLL load failed: The specified module could not be found.

• However, if I put the CRT assembly in the same folder as _socket.pyd:
  ImportError: DLL load failed: A dynamic link library (DLL)
  initialization routine failed.
  and I received the same popup dialog as above.

Putting the CRT assembly in the Apache bin folder or Apache modules
folder still gives the first error (module could not be found).

I apologize for the confusion; when I first experienced this problem I
tried to fix it by experimenting with putting the manifest in various
folders and I wasn't paying close enough attention to what error was
given when.

To summarize:
testpy.exe with CRT assembly in testpy.exe folder: init routine failed
and popup.
testpy.exe with CRT assembly in _socket.pyd folder: init routine failed
and popup.
testpy.exe otherwise: module could not be found

Apache with CRT assembly in _socket.pyd folder: init routine failed and
popup.
Apache otherwise: module could not be found

Mark Hammond <mhammond@users.sourceforge.net> added the comment:

I've no time to dig deeper now as I suspect testing will require removal
of the vc9 assembly from the GAC and testing with a local one, but some
comments:

Isn't the GAC just for .NET assemblies while the SxS cache is for non
.NET assemblies? Microsoft's naming schema is confusing as usually. ;)

Christian

I took a look at this with the debugger, as Mark recommended. The CRT's
DLLMain is called _CRTDLL_INIT, that in turn calls __CRTDLL_INIT.
__CRTDLL_INIT calls another function, _check_manifest.

_check_manifest calls an SxS function called FindActCtxSectionString.
It's looking for a string called "msvcr90.dll" in the
ACTIVATION_CONTEXT_SECTION_DLL_REDIRECTION section of the process's
activation context (activation contexts are data structures used by
SxS). That call fails (_check_manifest doesn't call GetLastError or
anything, it just returns FALSE right there). There's a comment there
that says:

/* no activation context used to load CRT DLL, means no manifest present
in the process */

What's bizarre is that python26.dll successfully loaded msvcr90.dll (the
global one from WinSxS), so it must have passed _check_manifest. It
seems like the activation context consists of the DLL's manifest
(_socket.pyd's in this case) and the exe's manifest, but no other ones,
regardless of what other libraries have been loaded. The documentation
doesn't seem to explain the interaction between manifests in different
modules. It's also annoying that this restriction (the _check_manifest
call) is completely artificial.

I don't know if any of this information is useful (I'm only
superficially familiar with activation contexts and the functions to
manipulate them). The code is all in crtlib.c (in the CRT source) if
someone else wants to take a look.

I've hacked together something that fixes the problem. I'm working on
making it a real patch, but the basis is:

• In DllMain (dl_nt.c), we call:
  case DLL_PROCESS_ATTACH:
  GetCurrentActCtx(&PyWin_DLLhActivationContext);
  AddRefActCtx(PyWin_DLLhActivationContext);
  case DLL_PROCESS_DETACH:
  ReleaseActCtx(PyWin_DLLhActivationContext);

  (where PyWin_DLLhActivationContext is a module level HANDLE global)

• The LoadLibraryEx (in dynload_win.c) changes to:

  ActivateActCtx(PyWin_DLLhActivationContext, &cookie);
  /* XXX This call doesn't exist in Windows CE */
  hDLL = LoadLibraryEx(... as currently ...);
  DeactivateActCtx(0, cookie);

This makes the "import socket" test case work for me. I'm not sure what
the impact will be should the .pyd file reference other assemblies.
Also, the calls above all need to move to function pointer calls as the
functions only exist on Vista and later (hence the slight delay in
making a real patch)

Attaching a patch which works for me against python 2.6. Only ever
tested on Vista (ie, where the function pointers etc all load)

Attaching a new patch with some typos in the comments corrected. While
I'm here, I also meant to mention (again!):

• No need to remove the the assembly from the sxs cache - the test
  "fails correctly" with VS2009 installed. FWIW, I compiled test.c from
  the command line with:
  % cl /I {pypath}\include /I {pypath}\PC /DDEBUG /D_DEBUG /MTd test.c
  /LINK {pypath}\pcbuild\python26_d.lib

• It appears the sxscache and the "gac" are indeed different things,
  although at least one MS page refers to the sxscache as the "native
  assembly cache" <wink/sigh>

I haven't been able to try this patch myself yet, but I see a potential
problem: the "cookie" variable is declared as a DWORD, while
ActivateActCtx expects a ULONG_PTR. DWORD and ULONG_PTR are only the
same thing in 32-bit Windows.

Also, where are you seeing that these SxS functions are "Vista or
later"? My XP kernel32.dll has all of them. MSDN says they're XP or Vista.

The patch works fine on my system (32-bit XP). Also I verified in
Process Explorer that there's only one instance of msvcr90.dll loaded.

Here's an option, though unfortunately not a trivial one: use a private
build of the C runtime.

I'm not sure whether the license allows us to do so.

Uploading a corrected patch; some old docs I saw said DWORD, and I
obviously neglected to fix every spot I used that, and I've changed
'Vista' to 'XP' in the patch comments.

We encountered a problem that Windows installer created with 2.5.2 could
not be installed with 2.6.1 as executing the post-installer script
fails. More details about the reason can be found from
http://code.google.com/p/robotframework/issues/detail?id=196. Is the
problem related to this issue or should I create new issue about that. I
can also provide simple project that could be used to test this out.

Is the problem related to this issue or should I create new issue about that.

You should assume that it is not related.

I can also provide simple project that could be used to test this out.

That would be necessary.

Over 3 years ago I wrote a plugin for the Xplane Flight Simulator.
This uses a SDK that Ben Supnik and myself created 6 years ago.
Our plugins are DLL's that our plugin manager DLL loads at run time.

The plugin embeds python and allows python scripts to be run from within
Xplane.
This has worked fine with python 2.3.x, 2.4.x, 2.5.x and 2.6.0.
I see the same problem as described here when using 2.6.1 with the
ctypes, socket libraries etc.
This also affects PyOpenGL as it uses ctypes.

I applied the patch that Mark posted and rebuilt python26.dll.
I tested this on Windows XP and Windows 7 Beta.
With the patched python26.dll my scripts now load properly.
The script that uses PyOpenGL also works fine now.

I thought that this feedback may be useful.

Mark, the patch is fine, please apply. Great work!

ack - I mis-clicked and accidentally removed message78811 and can't see
how to reinstate it. The message isn't critical, but I'm sorry about that!

Checked in to the trunk as as r69038 and svnmerge'd:

• release26-maint as 69040
• py3k as 69041
• release30-maint as 69043

After some Googling, I found a possible solution:

http://lists.wxwidgets.org/pipermail/wxpython-users/2008-November/081981.html

For some reason this patch is disabled with #ifdefs when Py_ENABLE_SHARED is not defined.

I'm compiling python as a static lib and have unresolved externals for both these functions: _Py_ActivateActCtx, _Py_DeactivateActCtx