Yep, this is getting a little hairy, but nothing we couldn't handle :-)
I did a little more research. Basically, this started with Graham's patch
that addressed a problem with modules being reimported (or something).
From Graham's message:
The basic problem revolves around the Python dictionary used to hold the
set of interpreters. The code in mod_python.c is trying to use the
Python GIL to provide exclusive access to that dictionary and any
subsequent creation of an interpreter.
The only catch is that in creating a new interpreter, the Python core
is, in someway I don't understand, swapping thread states at some point
which is allowing other threads to acquire the GIL.
So what Graham's patch does is create an APR lock (interpreters_lock) and
wrap all the access to the dictionary with calls to apr_mutex_lock/unlock.
I think the _real_ way to address this issue is to first find what is the
problem with using the Python GIL to serialize access to the interpreters
dictionary. Is this a Python bug, or are we not understanding GIL and
using it improperly?
BUT, given that the above question may be complicated to answer, and that
Graham's patch resolves the issue, another thought:
If the APR lock works, what is the point of using the GIL in addition?
Should we just use the APR-based lock alone? I.e., where we had (after
Graham's patch):
#ifdef WITH_THREAD
apr_thread_mutex_lock(interpreters_lock);
PyEval_AcquireLock();
#endif
we would use:
#ifdef APR_HAS_THREAD
apr_thread_mutex_lock(interpreters_lock);
#endif
_without_ a call to PyEval_AcquireLock() at all.
It should compile OK, and on platforms where APR has no thread support,
like you said, it's not an issue since no separate interpreters run in one
process at the same time.
Grisha
On Mon, 12 Sep 2005, Nicolas Lehuen wrote:
Duh, this is becoming difficult :)
I was thinking that if APR_HAS_THREADS was 0, then Apache was forcibly ran
in prefork mode, so there was no need for thread safety at all, given the
fact that mod_python would only run one interpreter thread. So if
WITH_THREAD was not defined, ORAPR_HAS_THREADS was 0, then we would not need
any thread safety code. Hence the definition of
MOD_PYTHON_WITH_THREAD_SUPPORT.
You're right in writing that a user could launch a new thread in Python,
provided that WITH_THREAD is defined, even if APR_HAS_THREADS==0. However,
having a look at the parts of mod_python.c where the thread safety was put
in, I think we can safely say that those parts are only called by mod_python
(through python_handler, python_cleanup etc who call get_interpreter). Those
parts are therefore always called in the same thread (if APR_HAS_THREADS==0,
that is) and there is no need for thread synchronization to be done (no
shared data between the main thread and the other user threads, no need to
release the GIL etc.).
BUT, I could be very, very wrong here, and your idea of reverting to a
conservative "shield python threading calls with WITH_THREAD and apr
threading code with APR_HAS_THREADS" is way more attractive to my tired mind
right now. So if you want I can revert all this
MOD_PYTHON_WITH_THREAD_SUPPORT hack.
Regards,
Nicolas
2005/9/12, Gregory (Grisha) Trubetskoy <[EMAIL PROTECTED]>:
I'm not sure I understand this, perhaps someone could write a message to
the list explaining what we're doing here so there is a record. Sorry if
I'm being slow-headed here.
To me it seems that when you use thread-related calls from Python, you
wrap those in Python defines (WITH_THREAD) and when you use thread-related
calls from APR, you wrap those in APR defines (APR_HAS_THREAD), and that's
all?
In other words - what does MOD_PYTHON_WITH_THREAD_SUPPORT accomplish that
the above does not.
Also, given:
#if(defined(WITH_THREAD) && APR_HAS_THREADS)
#define MOD_PYTHON_WITH_THREAD_SUPPORT 1
#else
#define MOD_PYTHON_WITH_THREAD_SUPPORT 0
#endif
Does this mean that if Python is compiled with thread support and APR is
not, MOD_PYTHON_WITH_THREAD_SUPPORT is 0 which means that the thread
safety code isn't there, but you still _can_ create threads because Python
will let you - isn't this asking for a segfault/deadlock/whatever?
Grisha
On Mon, 12 Sep 2005, Jim Gallacher wrote:
Gregory (Grisha) Trubetskoy wrote:
Shouldn't that be PYTHON_WITH_THREAD rather than
MOD_PYTHON_WITH_THREAD?
I understand it to mean that we want the thread handling code compiled
into
mod_python.
Compiling and testing right now.
Jim
On Mon, 12 Sep 2005, Nicolas Lehuen wrote:
I've checked in a changeset wherein I define
MOD_PYTHON_WITH_THREAD_SUPPORT
and use it everywhere WITH_THREAD was previously used. This should do
the
trick ! Now if someone (like Jim) can give us his +1, that would be
great.
Regards,
Nicolas
2005/9/12, Gregory (Grisha) Trubetskoy <[EMAIL PROTECTED]>:
Just wanted to add to this message that if Jim's version runs and
tests
with the trick below (envvars is executed prior to apache start, but
I
don't think the tests use it, so you'll probably just have to set
this
var
in the shell in which the tests are run), then this would be a
solution
for all FreeBSD issues and we could roll a beta 3 which will have a
great
change of being publicly released.
Grisha
On Mon, 12 Sep 2005, Gregory (Grisha) Trubetskoy wrote:
OK, found it. This should work on FreeBSD where Python is threaded
and
Apache
is not.
[snip]
And, if you built apache without thread support, you may need to add
the
following lines to $PREFIX/sbin/envvars:
LD_PRELOAD=/usr/lib/libc_r.so
export LD_PRELOAD
[snip]
On Mon, 12 Sep 2005, Gregory (Grisha) Trubetskoy wrote:
On Mon, 12 Sep 2005, Jim Gallacher wrote:
*** Warning: Linking the shared library mod_python.la against
the
*** static library
/usr/local/lib/python2.4/config/libpython2.4.a is
not
portable!
I think this was always there and its pretty harmless.
On qemu:
Syntax error on line 44 of
/usr/home/jim/tmp/mod_python/test/conf/test.conf:
Cannot load /usr/home/jim/tmp/mod_python/src/mod_python.so into
server:
/usr/home/jim/tmp/mod_python/src/mod_python.so: Undefined symbol
"pthread_attr_init"
This is because FreeBSD's libc comes in two versions - threaded
and
non-threaded. If Python is linked against the threaded ones and
Apache
against the non-thrreaded, then you get this problem. There is a
simple
fix for this - you just cause Apache to start with threaded libs,
but I
can't find any references to it right now and have to run off to a
meeting.
Grisha
It is quite possible I don't have things configured correctly on
the
QEMU version and hence the different undefined symbol but it
doesn't
really matter since it fails either way. I don't have time to
investigate further right now. I'll revisit this tonight.
Regards,
Jim
Regards,
Nicolas
#if APR_HAS_THREADS && defined(WITH_THREAD)
2005/9/11, Jim Gallacher <[EMAIL PROTECTED]
<mailto:[EMAIL PROTECTED]>>:
FYI, I found the following note in the INSTALL file in the
apache
source:
* If you are building on FreeBSD, be aware that threads will
be disabled and the prefork MPM will be used by default,
as threads do not work well with Apache on FreeBSD. If
you wish to try a threaded Apache on FreeBSD anyway, use
"./configure --enable-threads".
I'm also setting up FreeBSD under QEMU... so far so good, but
installing
anything using ports is really slow. QEMU's performance here
is
just
killing me. I guess I should have read the manual first and
used
the
binary packages for the software I wanted to install. :-(
Regards,
Jim
Jim Gallacher wrote:
Nicolas Lehuen wrote:
OK, I've checked in a version that compiles both on at
least
Win32 and
FreeBSD. I'm just testing if APR_HAS_THREAD is defined and
only
include the apr_thread_mutex_lock and unlock calls if it
is
defined.
Compiles a passes unit tests on Linux Debian sid with
mpm-prefork.
Now, on minotaur, APR_HAS_THREAD is defined as 0. Does
this
mean
that
Apache is not configured for threading ? Can we assume
that we
are in
the prefork model if APR_HAS_THREAD==0, so that we can
skip
all the
locking code ? Because that's what we do right now.
On Debian sid with apache2.0.54 mpm-prefork, APR_HAS_THREAD
==
1.
Jim
Regards,
Nicolas
2005/9/11, Nicolas Lehuen <[EMAIL PROTECTED]
<mailto:[EMAIL PROTECTED]>
<mailto: [EMAIL PROTECTED]
<mailto:[EMAIL PROTECTED]>>>:
Yes, this new code is something I commited on the
29/12/2004
(I used
the "blame" function of TortoiseSVN for that). It was a
patch by
Graham to fix MODPYTHON-2
<http://issues.apache.org/jira/browse/MODPYTHON-2>.
The problem is not in the patch, but rather in the fact
that
APR
seems configured without the thread support while Python
is
configured with thread support. mod_python.c assumes that
is
WITH_THREAD is defined, then the APR mutex functions are
available,
which is wrong. Maybe we should test for APR_HAS_THREADS
instead ?
In that case, won't this cause any problems on threaded
platforms ?
I don't know if this is a problem specific to minotaur or
to
all
version of FreeBSD. I'm currently downloading the ISOs of
FreeBSD
and I'll try using QEMU to run a FreeBSD setup on my
computer, but
that will be long and troublesome. If someone has more
clue
on this
issue, feel free to tell us :).
Regards,
Nicolas
2005/9/10, Jim Gallacher <[EMAIL PROTECTED]
<mailto:[EMAIL PROTECTED]>
<mailto:[EMAIL PROTECTED]
<mailto:[EMAIL PROTECTED]>>>:
I'm stubling around in the dark here, but maybe this
will
create a
spark
of an idea. I took a diff of mod_python.c from 3.1.4 and
3.2.1b and
isolated the lines which correspond to the compilation
error.
Compiler messages
-----------------
mod_python.c:34: error: syntax error before '*' token
mod_python.c:34: warning: type defaults to `int' in
declaration of
`interpreters_lock'
mod_python.c:34: warning: data definition has no type or
storage class
mod_python.c: In function `get_interpreter':
mod_python.c:131: warning: implicit declaration of
function
`apr_thread_mutex_lock'
mod_python.c:161: warning: implicit declaration of
function
`apr_thread_mutex_unlock'
mod_python.c: In function `python_init':
mod_python.c:517: warning: implicit declaration of
function
`apr_thread_mutex_create'
mod_python.c:517: error: `APR_THREAD_MUTEX_UNNESTED'
undeclared (first
use in this function)
Diff output
-----------
I've only copied the diff chunks which correspond to the
complier
errors
mentioned above.
--- mod_python-3.1.4/src/mod_python.c Sat Jan 29
13:25:28
2005
+++ mod_python-3.2.1b/src/mod_python.c Tue Sep 6
17:11:03
2005
@@ -31,6 +31,8 @@
* (In a Python dictionary) */
static PyObject * interpreters = NULL;
+static apr_thread_mutex_t* interpreters_lock = 0;
+
apr_pool_t *child_init_pool = NULL;
... snip ...
@@ -124,11 +128,15 @@
name = MAIN_INTERPRETER;
#ifdef WITH_THREAD
+ apr_thread_mutex_lock(interpreters_lock);
PyEval_AcquireLock();
#endif
... snip ...
@@ -149,6 +158,7 @@
#ifdef WITH_THREAD
PyEval_ReleaseLock();
+ apr_thread_mutex_unlock(interpreters_lock);
#endif
... snip ...
@@ -490,13 +506,15 @@
}
/* initialize global Python interpreter if necessary */
- if (! Py_IsInitialized())
+ if (initialized == 0 || !Py_IsInitialized())
{
-
+ initialized = 1;
+
/* initialze the interpreter */
Py_Initialize();
#ifdef WITH_THREAD
+
apr_thread_mutex_create(&interpreters_lock,APR_THREAD_MUTEX_UNNESTED,p);
/* create and acquire the interpreter lock */
PyEval_InitThreads();
#endif
So it would seem that the code causing the compile
problems
is new
for 3.2.
I also notice that in apr_arch_thread_mutex.h the
typedef
for
apr_thread_mutex_t is wrapped by #if APR_HAS_THREADS /
#endif.
Looking at the apache source in
srclib/apr/locks/unix/thread_mutex.c,
everything is also enclosed by #if APR_HAS_THREADS /
#endif.
eg, apr_thread_mutex_create, apr_thread_mutex_lock and
apr_thread_mutex_unlock.
Hopefully this will give someone a clue as to what may
be
going on
here
with FreeBSD.
Regards,
Jim
