[Catalyst-dev] Yet another IE crash

[hkclark at gmail.com]

On 10/25/06, hkclark at gmail.com <hkclark at gmail.com> wrote:
> On 10/25/06, Andy Grundman <andy at hybridized.org> wrote:
> > After fixing a few test server crashes caused by IE, I am getting one
> > that I can't figure out.
> >
>
> Whoa... major TCP RSTs going on here.  They are being sent with
> increasing frequency during the trace from the client-side of the
> connection... I obviously need to study the trace in more detail, but
> I wonder if the dev server has an aversion to mega-resets??
>

Here is an update for the group on my efforts re this issue in case
this triggers any ideas:

* The problem appears to only exist with the image files
(catalyst_logo.png in Andy's example).  The actual HTML text pages
work fine with no "RESET Explosion".

* I'm doing my testing of the development server using
'script/myapp_server.pl' with no options (if you want to try this at
home, just do "catalyst.pl MyApp;cd MyApp;script/myapp_server.pl" and
then point a copy of IE at your dev server and hit Ctrl-R quickly a
number of times).  (Note: I'm using IE 6.0 on Windows XP SP2.)

* A number of folks have pointed out articles talking about how IE can
be a little "TCP RST-happy".  However, I don't see the resets when I
point IE to the same webapp running until Apache with FastCGI... so
the question is, why does the dev server cause IE to generate so many
RSTs??  (One could argue that the dev server shouldn't die because of
too many RSTs, but I guess I'm more focused on what caused the RSTs in
the first place.)

* I'm wondering if it's related to HTTP 1.0 vs. HTTP 1.1.  As noted my
my post yesterday, IE works fine to the development server if you run
it through an SSH tunnel.  A tcpdump shows that this causes IE to drop
back to HTTP 1.0... I tried the Advanced Option in IE to disable HTTP
1.1 to see if that cause it to work with the dev server directly
(without the SSH port forwarding), but it seemed to still do HTTP 1.1
for the image files even with the checkbox unchecked. :-(

* I tried comparing traces of the exact same app between the dev
server and Apache with FastCGI.  Unfortunately, the traces are just
too different to help narrow things down.  For example, Apache keeps
giving "304 Not Modified" responses (I tried to get it to disable
doing 304s, but couldn't find the magic combo).

* I was getting strange decodes using wireshark (ethereal) where first
HTTP response header to the "GET /static/images/catalyst_logo.png" had
the first character chopped off (specifically, "Connection: close\r\n"
was being shown as "onnection: close\r\n").  I loaded my traces in a
copy of Sniffer Pro and did not see this and could not find other
evidence of this problem, so I have chalked it up to being a bug in
WireShark (I'm on v0.99.3).

So, all-in-all, I don't have a ton to show for my time on this... I
guess I'm hoping that the above info will cause a lightbulb to go off
for someone.

Regards,
Kennedy