This is a cut 'n paste job from an email (<022501c1c529$f63a9550$7f00000a@KOJ>) and only reformatted for better readability. It's not up to date but may be a good start for further research.
There are three basic filter types (each of these is actually broken down into two categories, but that comes later).
CONNECTION
AP_FTYPE_CONNECTION
, AP_FTYPE_NETWORK
)PROTOCOL
AP_FTYPE_PROTOCOL
,
AP_FTYPE_TRANSCODE
)RESOURCE
PROTOCOL
, but internal redirects and sub-requests can change
the content without ending the request. (AP_FTYPE_RESOURCE
,
AP_FTYPE_CONTENT_SET
)It is important to make the distinction between a protocol and a resource filter. A resource filter is tied to a specific resource, it may also be tied to header information, but the main binding is to a resource. If you are writing a filter and you want to know if it is resource or protocol, the correct question to ask is: "Can this filter be removed if the request is redirected to a different resource?" If the answer is yes, then it is a resource filter. If it is no, then it is most likely a protocol or connection filter. I won't go into connection filters, because they seem to be well understood. With this definition, a few examples might help:
The further breakdown of each category into two more filter types is
strictly for ordering. We could remove it, and only allow for one
filter type, but the order would tend to be wrong, and we would need to
hack things to make it work. Currently, the RESOURCE
filters
only have one filter type, but that should change.
This is actually rather simple in theory, but the code is
complex. First of all, it is important that everybody realize that
there are three filter lists for each request, but they are all
concatenated together. So, the first list is
r->output_filters
, then r->proto_output_filters
,
and finally r->connection->output_filters
. These correspond
to the RESOURCE
, PROTOCOL
, and
CONNECTION
filters respectively. The problem previously, was
that we used a singly linked list to create the filter stack, and we
started from the "correct" location. This means that if I had a
RESOURCE
filter on the stack, and I added a
CONNECTION
filter, the CONNECTION
filter would
be ignored. This should make sense, because we would insert the connection
filter at the top of the c->output_filters
list, but the end
of r->output_filters
pointed to the filter that used to be
at the front of c->output_filters
. This is obviously wrong.
The new insertion code uses a doubly linked list. This has the advantage
that we never lose a filter that has been inserted. Unfortunately, it comes
with a separate set of headaches.
The problem is that we have two different cases were we use subrequests. The first is to insert more data into a response. The second is to replace the existing response with an internal redirect. These are two different cases and need to be treated as such.
In the first case, we are creating the subrequest from within a handler
or filter. This means that the next filter should be passed to
make_sub_request
function, and the last resource filter in the
sub-request will point to the next filter in the main request. This
makes sense, because the sub-request's data needs to flow through the
same set of filters as the main request. A graphical representation
might help:
Default_handler --> includes_filter --> byterange --> ...
If the includes filter creates a sub request, then we don't want the data from that sub-request to go through the includes filter, because it might not be SSI data. So, the subrequest adds the following:
Default_handler --> includes_filter -/-> byterange --> ... / Default_handler --> sub_request_core
What happens if the subrequest is SSI data? Well, that's easy, the
includes_filter
is a resource filter, so it will be added to
the sub request in between the Default_handler
and the
sub_request_core
filter.
The second case for sub-requests is when one sub-request is going to
become the real request. This happens whenever a sub-request is created
outside of a handler or filter, and NULL is passed as the next filter to
the make_sub_request
function.
In this case, the resource filters no longer make sense for the new request, because the resource has changed. So, instead of starting from scratch, we simply point the front of the resource filters for the sub-request to the front of the protocol filters for the old request. This means that we won't lose any of the protocol filters, neither will we try to send this data through a filter that shouldn't see it.
The problem is that we are using a doubly-linked list for our filter stacks now. But, you should notice that it is possible for two lists to intersect in this model. So, you do you handle the previous pointer? This is a very difficult question to answer, because there is no "right" answer, either method is equally valid. I looked at why we use the previous pointer. The only reason for it is to allow for easier addition of new servers. With that being said, the solution I chose was to make the previous pointer always stay on the original request.
This causes some more complex logic, but it works for all cases. My concern in having it move to the sub-request, is that for the more common case (where a sub-request is used to add data to a response), the main filter chain would be wrong. That didn't seem like a good idea to me.
The final topic. :-) Mod_Asis is a bit of a hack, but the
handler needs to remove all filters except for connection filters, and
send the data. If you are using mod_asis
, all other
bets are off.
The absolutely last point is that the reason this code was so hard to
get right, was because we had hacked so much to force it to work. I
wrote most of the hacks originally, so I am very much to blame.
However, now that the code is right, I have started to remove some
hacks. Most people should have seen that the reset_filters
和add_required_filters
functions are gone. Those inserted
protocol level filters for error conditions, in fact, both functions did
the same thing, one after the other, it was really strange. Because we
don't lose protocol filters for error cases any more, those hacks went away.
The HTTP_HEADER
, Content-length
, and
Byterange
filters are all added in the
insert_filters
phase, because if they were added earlier, we
had some interesting interactions. Now, those could all be moved to be
inserted with the HTTP_IN
, CORE
, and
CORE_IN
filters. That would make the code easier to
follow.