mod_fcgid 2.3.4 problem with uploading data > 64k

[adipose]

I have been able to confirm this bug but have not designed a test
case.  It should be fairly easy to duplicate, however.  My particular
test case was using phpicalendar 2.31 with publish.php on Sunbird
calendars.  Previously, this worked fine, but upon using fcgid 2.3.4,
I started getting calendar corruption (the publish is done by reading
the  php://input stream).

I was able to verify that the corruption occurred at right around the
64K mark of the uploaded data.  Basically it appeared that a chunk of
data was dropped at that point, but more data was written afterward.
Upon realizing this, I began to research any information having to do
with 64K.  I learned that there is a 64K "limitation" by default in
fcgid, via:

FCGIDMaxRequestInMem

The default value of this variable is 64K.  Upon changing this to a
larger value (I used 2^21, since that matches the php upload limit
default), the corruption disappeared.

Obviously, since the value of the variable only is supposed to change
WHERE the data is buffered, and not WHAT is buffered, there must be
some sort of bug, here.  As soon as there is a requirement to cache on
disk, the data is not correctly fed to the input stream.  Something is
lost.

I'm not sure the best way to make an easily reproducible test case,
since my configuration requires php.  Perhaps a perl cgi to read the
input, and a simple web request html file to post it...

-Dan

[Jeff Trawick]

On Tue, Nov 10, 2009 at 4:24 PM, Dan Hulme <dhulme@...> wrote:
> I have been able to confirm this bug but have not designed a test
> case.  It should be fairly easy to duplicate, however.  My particular
> test case was using phpicalendar 2.31 with publish.php on Sunbird
> calendars.  Previously, this worked fine, but upon using fcgid 2.3.4,
> I started getting calendar corruption (the publish is done by reading
> the  php://input stream).

Use this patch:

http://svn.apache.org/viewvc?view=revision&revision=826829

[adipose]

Aha, thank you!  This is the bug, indeed.

-Dan

On Tue, Nov 10, 2009 at 1:27 PM, Jeff Trawick <trawick@...> wrote:

> On Tue, Nov 10, 2009 at 4:24 PM, Dan Hulme <dhulme@...> wrote:
>> I have been able to confirm this bug but have not designed a test
>> case.  It should be fairly easy to duplicate, however.  My particular
>> test case was using phpicalendar 2.31 with publish.php on Sunbird
>> calendars.  Previously, this worked fine, but upon using fcgid 2.3.4,
>> I started getting calendar corruption (the publish is done by reading
>> the  php://input stream).
>
> Use this patch:
>
> http://svn.apache.org/viewvc?view=revision&revision=826829
>