httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ron Park <ronald.p...@cnet.com>
Subject Two mod_include problems
Date Wed, 09 Jul 2003 19:50:08 GMT
We have run across two problems in mod_include that were
corrupting our pages and would like to share our patches.

Both of these situations are highlighted by situations where
bucketeering and brigading are accentuated (mostly through
the use of mod_proxy to a remote host, but should be testable
with simple includes of appropriately sized (8K+) files).

The first problem was appearing as a few (important) chars
being truncated right before the occurrence of a virtual
include.  The situation was heightened by heavy use of
mod_proxy to retrieve these files and the smaller brigade
size that employs (about 1.5K).

After snooping the results of the proxy request, we verified
that the returned file was looking just fine with various
calls to <!--#include virtual... --> ... but after processing
by mod_include, some of the (valid) includes where failing.

In particular, we tracked down a case where the start of a
comment at the very end of the proceeding bucket would cause
a miscount of the start of the location of the include in
the current bucket.

As an example, consider a bucket that ends as follows:

... <!

and the new bucket starts as

-- some comment --><img src="a.gif"><!--#include ...

The problem we were seeing in the mod_include parsed page was
that the two chars proceeding this include were being eaten.
So, we would see the following string:

... <!-- some comment --><img src="a.gif<!-- missing include -->

As you might imagine, the missing end quote was troublesome. :)
(Not to mention the bad include...)

At line 438 of mod_include, we see the code is set up to handle
this 'left over partial match' and it increments the temporary
buffer pointer, c, by 2 as it walks past the '--' at the start
of this bucket.  So it (as far as I could determine) adds the
left over '<!' to this bucket and continues on, calling the
quick scanner function, bndm(), passing in the recently updated
parameter 'c'.  The result of this is 'pos', an offset into
the bucket supposedly representing the start of the include...
but it's off by 2!  Later, it proceeds to 'split' the bucket
at what it thinks is the start of the include (but is actually
the '">' from the proceeding tag) and fails to figure out what
include command is being requested, hence the 'missing' comment.

This patch which doesn't really fix the problem but corrects the
resulting problem... we hope someone with a better understanding
of mod_include and buckets can come up with a true solution.

======

The second problem actually occurs across brigades (we learned
an awful lot about buckets and brigades in diagnosing these
problems! :D)  And, instead of removing characters was instead
adding large extra chunks.

This problem revolves around the use of 'if' 'else' and 'elseif'.
Originally, we thought the problem was related to <!--#else -->
but later we found an troublesome <!--#if ... --> that lead to
us finding the real culprit.

As an example, consider the following example:

... <!--#if ... -->Some if text ... end if text <!--#endif -->

The problem reveals itself if the 'Some if text ...' is in one
brigade and the 'end if text' is in the next *and* the if's
condition is false (hence the if's conditional body should
*not* be present after mod_include works it's magic).

What we found is that we were getting 'Some if text ...' out
of mod_include... but not the 'end if text' when the condition
was false. :(

We finally tracked the problem down because of the comments
on line 3200 of mod_include.c:

        /* Inside a false conditional (if, elif, else), so toss it all... */
        if ((dptr != APR_BRIGADE_SENTINEL(*bb)) &&
            (!(ctx->flags & FLAG_PRINTING))) {
            apr_bucket *free_bucket;
            do {
                free_bucket = dptr;
                dptr = APR_BUCKET_NEXT (dptr);
                apr_bucket_delete(free_bucket);
            } while (dptr != APR_BRIGADE_SENTINEL(*bb));
        }

Rightly so, this loop makes sure it doesn't throw out the
brigade's sentinel.  Surely this would be a 'bad thing'.

However, earlier in the code (3015 ... 3044), we have this:

            if (tmp_dptr != NULL) {
                ...
            }
            else {
                 /* remainder of this brigade...    */
                dptr = APR_BRIGADE_SENTINEL(*bb);
            }

And tmp_dptr, is set via line 2937:

            tmp_dptr = find_start_sequence(dptr, ctx, *bb, &do_cleanup);
 
But since there's no start sequences in the if's conditional
text, tmp_dptr is set to NULL (left as an exercise for the
reader :).

So the code would get to line 3200... and dptr would *always*
be pointing at the sentinel. That just clearly seems wrong.
The do loop would never be executed and the <!--#if-->'s body
would never be thrown away.  Exactly what we were seeing. :(

Once again, we don't claim the patch really fixes the problem;
it just tries to correct the situation we were seeing in the
most minimal way.  In fact, we kept (what we think is) the 'bad'
code in case there's some strange scenario where the code really
is good.

Attached is a single patch to fix both problems.


Mime
View raw message