Graham,

The patch is faulty with the 3rd hunk getting rejected. It looks like you generated it from an earlier revision, but one of the lines in the hunk #3 was included in your commit for r380087. The offending line is:

+                      "python_filter: Can't get/create interpreter.");

Beyond that, the code looks fine. I'm not familiar with filters so I may not be the right person to say if it's correct, however your explanation makes sense, so I'm +1.

Since we are discussing Python*Filter, can someone explain why it is only allowed in the server config context, whereas SetInputFilter and related directives are allowed in any context? Is this a mod_python feature or a bug, or is it just the way filters work?

Jim

Graham Dumpleton (JIRA) wrote:
     [ http://issues.apache.org/jira/browse/MODPYTHON-112?page=all ]

Graham Dumpleton updated MODPYTHON-112:
---------------------------------------

    Attachment: grahamd_20060223_1.diff

Attached patch "grahamd_20060223_1.diff".

Would like confirmation on final solution for this issue before I commit 
changes.

Behaviour would be such that req.phase is always set to be the phase of the currently 
executing handler. If an input or output filter is triggered it will no longer overwrite 
the value of req.phase. Thus, if in the content handler, req.phase will stay as 
"PythonHandler" for the whole time the handler is executing.

If there is no mod_python content handler and just a mod_python input or output filter, 
accessing filter.req.phase from the filter will result in the Python None value.  This 
will be the case even if a prior phase to the content handler executed as a mod_python 
handler, eg. "PythonFixupHandler", as req.phase will be cleared at the end of 
any mod_python handler phase.

That is, at no time will req.phase be set to be either PythonInputFilter or 
PythonOutputFilter like it used to. Whether a filter is running in input or 
output mode is determinable from filter.is_input so no need for filter 
directive name to be put in req.phase. This does not cause any problems for 
CallBack.FilterDispatch() in apache.py as it didn't use req.phase, but used 
filter.is_input where necessary. The change to CallBack.HandlerDispatch() in 
apache.py to cache the phase the handler is executing in as described 
previously for this issue, is no longer necessary with the way this fix works.




If using filters value of req.phase only valid up till first 
req.read()/req.write().
------------------------------------------------------------------------------------

        Key: MODPYTHON-112
        URL: http://issues.apache.org/jira/browse/MODPYTHON-112
    Project: mod_python
       Type: Bug
 Components: core
   Versions: 3.1.4, 3.2
   Reporter: Graham Dumpleton
Attachments: grahamd_20060223_1.diff

The request object provides a member variable called 'phase' described as:
 The phase currently being being processed, e.g. "PythonHandler". (Read-Only)
If no Python based input and output filters are used, the value of req.phase 
will be constant for the life of a request phase. If however you use an input 
or output filter, the value of req.phase can change.
Consider that we are in the content handler phase and where there is a Python based output filter, 
but no Python based input filter. On initially entering the request handler, the value of req.phase 
will be "PythonHandler". As soon as req.write() is called however, the value of req.phase 
changes to "PythonOutputFilter".
Now, if there is a Python based input filter, but no Python based output filter, the 
value of req.phase will change to "PythonInputFilter" as soon as req.read() is 
called.
If there are both Python based input and output filters, the value of req.phase will in turn change 
to "PythonInputFilter" and "PythonOutputFilter" as req.read() and then 
req.write() are in turn called.
The reason for all this is that in the get_request_object() function of 
src/mod_python.c, it contains code:
   /* make a note of which phase we are in right now */
   Py_XDECREF(request_obj->phase);
   if (phase)
       request_obj->phase = PyString_FromString(phase);
   else
       request_obj->phase = PyString_FromString("");
That is, whenever called to get the request object, it will update req.phase. 
This will occur even if the request object had already been created, as will be 
the case when there is a Python based content handler and either a Python based 
input or output filter.
Overall this behaviour is a bit strange and unexpected. It would seem to me that it if there is both a 
handler and a filter, that the value of req.phase would be left as the name of the handler phase. Ie., it 
would stay for example as "PythonHandler" and not be changed to "PythonInputFilter" or 
"PythonOutFilter".
One can't just change the code in get_request_object() not to update it if 
already set, as it has to be updated when one moves from one phase to another. 
One has to contend with where this function is called in python_filter() 
function, namely:
   /* create/acquire request object */
   request_obj = get_request_object(req, interp_name,
                                    
is_input?"PythonInputFilter":"PythonOutputFilter");
First step may be simply not to pass in "PythonInputFilter" or 
"PythonOutputFilter" and instead just call it as:
   request_obj = get_request_object(req, interp_name,0);
At the same time change get_request_object() to:
   Py_XDECREF(request_obj->phase);
   if (phase)
       request_obj->phase = PyString_FromString(phase);
   else if (!request_obj->phase)
       request_obj->phase = PyString_FromString("");
Ie., result will be that if req.phase is set, leave it alone when called by 
python_filter() with phase set 0. If req.phase isn't already set, set it to the 
empty string.
The consequences of this are that if there are filters but no Python based handlers, then 
req.phase will get set to an empty string in that case. Another strange case is that if 
the only Python based handler was for an early phase than what is consuming or generating 
the content to be processed, then req.phase will be set to a value corresponding to that 
earlier phase and not where the current action is. For example, if there was an 
AccessHandler but then content came from a static file, output filter would see 
"AccessHandler".
Thus, whatever is done, there will be some strangeness. Thus, question remains 
of what should be done, or if it should be left that way and that documentation 
changed to say that req.phase is only valid up until first call to req.read() 
or req.write() within a handler.
This is not an ideal situation though as a handler may want to interogate 
req.phase after req.read() or req.write() has been called for some reason, 
which would yield incorrect results if a filter is being used. An example of 
where this occurs is in error reporting in the HandlerDispatch of 
mod_python.apache itself. When it is generating an error message, the phase 
shown in the error message will be wrong if there was a filter. It should 
perhaps at least be changed to save away req.phase at start of dispatch so it 
knows it is correct later for any error messages.
Any comments?????




Reply via email to