Date: Fri, 2 May 1997 16:57:51 -0700 (PDT)
From: Dean Gaudet <dgaudet@arctic.org>
To: johnh
Subject: http://gost.isi.edu/~johnh/SOFTWARE/APACHE/index.html
Message-Id: <Pine.LNX.3.95dg3.970502163926.19228o-100000@twinlark.arctic.org>
Mime-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII

Just wandering your pages again.  Here's some notes you might want to add
to the above link.  It's stuff that all server developers will probably
have to go through at some point. 

The short-initial-segment problem is definately fixed in Apache 1.2 as of
1.2b7.  Something like this: 

All existing flushes were removed.  When the server is reading a request
line (GET /foobar HTTP/x.y) it will flush output if select() indicates
that a read() would block. 

This was chosen as a solution that works with both pipelining and
non-pipelining persistant clients.  A non-pipelining client, and
pipelining clients without any outstanding requests both require the flush
at this point.  A pipelining client with outstanding requests doesn't need
the flush.  This solution also avoids any need to keep timers and flush
after a certain amount of time has passed. 

One thing to watch out for here is that clients send an extra CRLF after a
POST request because the CERN webserver require it.  So the server has to
eat all CRLFs that appear before a request.

The original reason earlier versions of Apache flushed after the http
response headers was to work around a browser bug involving keepalive. 
Probably in Navigator version 2, for which we now disable all keepalive in
the default configuration.

Subsequent to removing the flush we discovered another Navigator bug
present in all versions to at least 4.0b2.  If the trailing CRLF of the
http response header fall at the 256th or 257th byte of a "packet" then
navigator will misinterpret the headers (and corrupt the result).  It's
hard to say exactly what goes wrong, we guess their initial read of a new
request is into a 256 byte buffer and further reads are in a larger
buffer.  In any event apache detects this condition and adds a small
padding header to force the trailing CRLF to be further into the packet. 

Dean