[Python-Dev] Windows x64 & bsddb 4.4.20 woes

[Trent Nelson]

Ah, and to think I just fixed 4.4.20 ;-)

Removing the dependency on db_static.vcproj and merging the relevant source code files into _bsddb.vcproj did the trick -- all x64 bsddb-related tests now pass.  The only issue with this approach is that it locks _bsddb.vcproj into 4.4.20.  However, considering that this approach (i.e. bringing their source files into our build instead of linking against a static lib compiled with wildly incompatible flags) only took me about two minutes to implement and immediately fixed every bsddb problem I was encoutering, I'm convinced it's the right way to go.  (I can separate the dependencies easily enough.)

Woeful PyCon/hotel connectivity is preventing me from getting to bugs.python.org at the moment; I'll raise a ticket later to capture this stuff and we can move the discussion there once I've attached some patches.

    Trent.


From: Gregory P. Smith [greg at krypto.org]
Sent: 14 March 2008 00:23
To: Trent Nelson
Cc: python-dev at python.org; Jesus Cea
Subject: Re: Windows x64 & bsddb 4.4.20 woes


On 3/13/08, Trent Nelson <tnelson at onresolve.com> wrote:
Hey Greg,

I'm at PyCon indeed, staying through the sprints 'til next Thursday.  I'll drop you a note offline re catching up.

The other query I had was whether or not I should try a later version of BerkeleyDB -- are we committed to 4.4.20 (or 4.4.x?) for 2.6/3.0 or is it worth investigating newer versions?  Martin/Jesus, any thoughts on this?


Python 2.6/3.0 should be built on Windows using BerkeleyDB 4.5.x for now.  4.6.x is out but has some bugs on some platforms so i don't recommend shipping our release using it; 4.7.x is in beta and some bugs are being worked on; if its out and shows no signs of obvious issues before the 2.6/3.0 beta period is over I recommend we build our binary releases using it.  Otherwise 4.5 it will be.  There is no reason to use 4.4.x.



Regarding the db_static build and conflicting compile/link options -- I'm going to bring the db_static source directly into the _bsddb project (for now) which should make this a lot easier to debug.

    Trent.


From: Gregory P. Smith [greg at krypto.org]
Sent: 13 March 2008 22:00
To: Trent Nelson
Cc: python-dev at python.org; Jesus Cea
Subject: Re: Windows x64 & bsddb 4.4.20 woes



I haven't built the bsddb stuff on windows myself in a few years and have never had access to a windows x64 system so I'm no silver bullet.  Making the BerkeleyDB compile and link options match with those of python is the first place I'd start.  Also you should be able to make a debug build of BerkeleyDB (though it sounds like you may have tried that already?).  Next off in the debugging i'd take a look at the assembly to see what exactly it was failing to do.

If you're at PyCon right now we should meet up and try to figure it out (I just arrived).


On 3/13/08, Trent Nelson <tnelson at onresolve.com> wrote:
I've been trying to give the Windows x64 builds a bit of TLC the past few evenings.  I managed to get a successful build with all external modules last night (Tcl/Tk required about a half a dozen code/configuration changes each in order to build in a Windows x64 environment with Visual Studio 9, I'll deal with that in a separate thread or roundup issue).

Unfortunately, though, we're back to more bsddb issues.  I got about 15 tests in without error before test_whichdb ran, which results in the following being called in dbhash.py:

        return bsddb.hashopen(file, flag, mode)

I can trace that call to DBEnv_open() in _bsddb.c:

static PyObject*
DBEnv_open(DBEnvObject* self, PyObject* args)
{
    int err, flags=0, mode=0660;
    char *db_home;

    if (!PyArg_ParseTuple(args, "z|ii:open", &db_home, &flags, &mode))
        return NULL;

    CHECK_ENV_NOT_CLOSED(self);

    MYDB_BEGIN_ALLOW_THREADS;
    err = self->db_env->open(self->db_env, db_home, flags, mode);
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Placing a breakpoint at the line above and stepping in results in Visual Studio reporting: " A buffer overrun has occurred in python_d.exe which has corrupted the program's internal state. Press Break to debug the program or Continue to terminate the program.".  FWIW, the exception is being raised as part of the /GS buffer overflow checks (implemented in gs_result.c, which is provided in my VS9 installation).

This has been annoyingly awkward to debug.  I can't break down that call into multiple steps in order to try place breakpoints in the db_static module.  The callstack isn't that useful either:

_bsddb_d.pyd!__crt_debugger_hook()
_bsddb_d.pyd!__report_gsfailure(unsigned __int64 StackCookie=2211040)
_bsddb_d.pyd!__GSHandlerCheckCommon(void * EstablisherFrame=0x000000000021bce0, ...)
_bsddb_d.pyd!__GSHandlerCheck(_EXCEPTION_RECORD * ExceptionRecord=0x000000000021bbc0, ...)
ntdll.dll!00000000773ae13d()
[Frames below may be incorrect and/or missing, no symbols loaded for ntdll.dll]
ntdll.dll!00000000773aea57()
ntdll.dll!00000000773b59f8()
_bsddb_d.pyd!__os_strdup()  + 0x18 bytes
_bsddb_d.pyd!__os_tmpdir()  + 0x281 bytes

You'd think placing breakpoints in db 4.4.20's __os_strdup and __os_tmpdir methods would do something, but alas, the bufferoverflow exception is raised before any breakpoints are set.  This makes me suspect there's something funky going on with the entire build and linking of db_static (VS should honour those breakpoints if the code is being entered, I even added db_static to pcbuild.sln and rebuilt but no dice).  I've noticed that they're not using consistent compiler flags by default (we use /GS, they use /GS-, we allow function level linking, they don't -- note that I did change db_static's options to align with _bsddb's but the bufferoverflow exception is still being thrown).

Greg, Jesús, I'm CC'ing you guys as stfw'ing seems to bring back you two the most when it comes to bsddb issues.  I've still got a list of things to try with regarding to debugging this x64 issue, but I wanted to reach out now to see if anyone else had encountered it before.  Has bsddb ever been built successfully on Win64 and passed all tests or am I venturing into new ground?

Martin, you've changed externals/bsddb-4.4.20 with regards to x64 builds recently -- have you been able to get things working in your x64 environments?

Regards,


        Trent.