W3C home > Mailing lists > Public > ietf-http-wg@w3.org > January to March 2004

Re: Can the response entity be transmitted before all the request entity has been read?

From: Jamie Lokier <jamie@shareable.org>
Date: Tue, 16 Mar 2004 06:00:35 +0000
To: Alex Rousskov <rousskov@measurement-factory.com>
Cc: HTTP WG <ietf-http-wg@w3.org>
Message-ID: <20040316060035.GA13637@mail.shareable.org>

Thanks for your detailed response.

Alex Rousskov wrote:
> >      write "Status: 200\n\n";
> >      write "Thank you for your submission.\n";
> >
> >      while ($x = read (...)) { store ($x) }
> >
> >      write "Great, all done.\n";
> While it is easy to imagine such an application, tasking a proxy to
> "rewrite" or "fix" application logic to fit HTTP restrictions seems
> like a bad idea. IMO, upon receiving "Status: 200\n\n" and sending off
> response headers, your proxy should become a "tunnel" and not try to
> second-guess the application intent.

Part of the mixed message here is that I'm simultaneously writing a
server which is intended to be well written, robust, protocol
compliant, persistent when possible, no deadlocks, and so forth; and
I'm also writing an application or two.  The applications are actually
the motivation for the server.

So the question of whether the above kind of application is worth
supporting is important, because I intend to try it out.

By the way, pure tunnelling leads to deadlock: the application can get
stuck writing if the client isn't reading the response until it
transmits all the request, and all the TCP windows fill up.  I don't
like that deadlock because it isn't necessary and it is practical to
eliminate it in the server.  It's messier to eliminate it in the
application, and anyway why do it once per application instead of once
in the server?

You're thinking: out of memory.  Actually no.  To avoid deadlock, this
is what I do: If writing would block, and there is data available to
read, read it.  The buffer holds _read_ data, and is therefore limited
by the maximum permitted request entity size.  The maximum is asserted
both for content-length and chunked requests.  The buffer is required
*somewhere*, either in the server or in the application itself, or in
backing storage for them, so there is no added resource consumption
from this technique when it's implemented properly.

> > I'd rather put the hard requirements, and every feature that will
> > help with robustness, in the _server_, rather than document it as a
> > requirement that applications have to follow.  It's the server's job
> > to keep the communication good as reliable as possible, insulating
> > the application.
> IMO, the "Do No Harm" rule trumps the "Try to change the world to the
> better" rule, especially for proxies (which is what you are
> implementing in this context). If you can reliably convert garbage
> into compliant output, do so. If your smart conversion algorithm
> silently breaks a few innocent applications, then do no smart
> conversion.

:) Well, you'll be glad to know that for now, especially as I want to
try it out, the server allows the application to read and write as it

It's not entirely a tunnel: as said above, the server takes care of
deadlock avoidance for clients which aren't reading until they've
finished writing.  That simplifies the app code, without changing
anything unless that situation arises and the deadlock would really

Obviously the server is in change of chunking, (de)compression,
boundary checks against content-length etc.  I'll take that as understood.

> If you need a negative example, consider Apache 2.x problems with
> smart content-length guessing algorithm that, AFAIK, still stalls a
> few simple CGIs that work fine with Apache 1.x:
> http://nagoya.apache.org/bugzilla/show_bug.cgi?id=23528

Yeah, that's a bug in Apache.  It's not at fault for being smart, it's
that it's code simply has a bug.  The idea is good as it allows some
HTTP/1.0 connections to be persistent when they otherwise wouldn't be.

> Specifically, if your server will be able to detect and reject
> applications that write before reading, fine. If your server delays
> any application output until it thinks there is no more input, your
> feature is probably going to be a popular target for denial of service
> attacks (for example) and you are probably going to deadlock
> applications that write more than one buffer worth of data (another
> example).

Presently I am allowing the app to write before it's finished reading,
because I'd like to try out that capability.  I might add a
per-request option to buffer all the input before writing anything -
it's fits easily with the deadlock-avoidance buffering, and could be
useful to some applications.

If I did restrict writing until reading was complete, the DOS attack
would be no different than the DOS where people try to send maximum
size request entities, e.g. uploading lots of large files.  That's the
only effect on the buffering algorithm: it ends up buffering up to one
maximum size request entity per request, and the overall memory and
disk management for that can certainly be constrained.

Note that the response doesn't get buffered infinitely.  Response
generation is blocked while the request is being buffered up for
whatever reason.

It's not possible to omit the buffering *somewhere*: for clients which
send a whole request before reading the response, the entire request
has to be buffered or stored *somewhere*, either in the server or in
the application, to resolve the deadlock.

> It is possible to have an HTTP-to-Applications API with enough logic
> and controls that optimizations you mention are very appropriate and
> safe. Is CGI such an interface? Do CGI specs document these things?

There is a CGI spec draft (1.2 at the moment), but it is not so well
written as HTTP/1.1.  Not to mention that quite a few of the CGI
meta-variables are implemented in various ways, and even the spec'd
ones miss out important info, so everyone adds a few more non-standard
variables.  (Like REQUEST_URI).

But that's beside the point, I'm not writing a CGI interface, I'm
writing and designing an HTTP-to-Application API with enough logic
etc. to do as you suggest.  At the same time, trying to keep it as
simple as possible - only the necessary controls.

> > That's contrary to most server implementations: they do give the
> > application control over when to read and write, which is the
> > opposite of what you're suggesting here.
> Am I? A tunnel is exactly what gives the application unlimited ability
> to read and write at any time, at will.

Exactly.  Too much control over the HTTP, like you said to avoid ;)
(Facetious comment; please ignore ;)

> The latter [100 Continue for PUT and POST] is a pro-active behavior
> intended to help RFC 2068 agents.  Unfortunately, it requires
> compliant RFC 2616 support for 100 Continue in proxies. My bet that
> sending 100 Continue pro-actively will hurt in more cases than it
> will help, but I have no data to prove that.

Hmm.  Maybe follow that SHOULD only if Via isn't defined?

If the client delay is small I'm not bothered.  Do you have any data
on how long those RFC 2068 agents will delay sending the request

> Moreover, there was a paper that formally proved that 100 Continue
> leads to deadlocks in certain compliant environments, so we are
> probably talking about a partially broken mechanism here anyway.

Hmm.  I'd like real data on what to do here.  If you can find the
paper or any other info, that would be very helpful.

I don't see any deadlock scenarios with the way I have implemented it.
Perhaps the deadlock occurs when it's coded in a different way (I've
been careful, most http implementors aren't half as cautious, judging
by the code I've read).  Or maybe I just have yet to see it.

I will have a peek at Apache's code to see what it does for the RFC
2068 clients. Squid won't give reliable answers: it's HTTP/1.1 support
is still in the development phase.

(I've looked at enough implementations to find loads of other quirks;
may as well see if anyone put in a comment about this).


Hmm.  Apache 2 doesn't satisfy the MUST of 8.2.3.  It won't sent
100-continue if the request has Content-Length: 0.

Both Apache 1.3.29 and 2.0.48 (current versions) apply the rule that
Expect: 100-continue with HTTP-Version >= 1 causes 100-continue to be
sent.  Neither of them apply the rule for supporting RFC 2068 clients.

Apache 1.3.29 won't send it with error responses, according to a quick
skim of the code, but a couple of servers I poked at didn't behave
like maybe I misunderstood something, or those servers are configured
to be more complicated.

A quick try at www.microsoft.com :) reveals IIS/6.0 in their setup
sends 100-continue to a POST with Expect: 100-continue, despite the
404 response.  More checking: it sends 100-continue to a POST without
Expect: 100-continue, despite the 404 response.

phttpd-1.10.4, thttpd-2.25b and lighttpd-1.0.3 don't send it ever,
despite all of them claiming to offset some level of HTTP/1.1.  I
guess they predate RFC 2616.

> > That's a crucial question.  Should I either enforce that in the
> > server, by insisting on reading the whole request before allowing
> > any text of a non-error response to be sent (error can be sent
> > immediately), or document it as an application requirement: that the
> > application must do all its reading before it writes anything?
> You cannot enforce this at the server without deadlocking or killing
> the application or running out of server memory. Imagine an
> application that writes more than you can buffer before it reads
> anything.

No.  The application can't do that: it will block when writing blocks.
The server's large buffer is the size of a maximum _request_ entity,
and that storage is unavoidable one way or another.  Everything else
is limited to an appropriate I/O block size.

> > It seems that existing servers, e.g. Apache, thttpd and all the
> > others don't do either: they allow request to be read by the
> > application when it likes, the response to be sent before all the
> > request is read if the application likes, and don't document this as
> > a problem.  It's for application writers to be aware of it.
> That's what I would do too, as far as code is concerned. Documenting
> potential problems is always good, of course, especially if you can
> give specific real-world examples.

I will document it as well.  If there are clients which accept the
non-error response before all the request is read, that is certainly a
feature worth letting the app have access to.

The reason I asked all this is that if practically all real clients
fail with it (I haven't reached the stage of testing yet), then the
server may as well constrain the app, probably by complaining at it.

> > I'd simply like to know whether it's best to program the server to
> > enforce that, knowing it's a common/rare client weakness, or to not
> > enforce it but recommend it in the application interface
> > documentation, or to permit it if it actually works in practice.
> Make your HTTP-to-application proxy as simple as possible. Warn of
> possible problems if the tunnel interface is abused. Let applications
> decide how they want to deal with those problems.

Ok.  Now give me advice as application author: I'd like to know
whether it a common or rare client weakness, so whether I should
consider using that technique or not.

If there's only some well known, old clients which don't like it, then
a match on the User-Agent string which activates the request entity
buffer will be the right thing to do.  The server can include it in
the plethora of other User-Agent quirks it already works around.  (I'd
rather put such knowledge in the server which already has that buffer
for deadlock avoidance, than in N applications).

If, however, lots of clients don't like it, then it isn't worth using
the technique at all, and it would be better for the server to
complain when the app erroneously tries it -- even if it's just a
warning which can be disabled.

> > My strategy is to copy Apache's well-tested "lingering close":
> > shutdown(fd,SHUT_WR) followed by reading everything for up to 30
> > seconds, or until 2 seconds passes with no incoming data, then the
> > full close().
> Cool. I hope this well-tested algorithm is not what breaks CGIs in
> Apache 2 :-).

The algorithm has been in Apache 1 for quite a while.

But maybe  Apache 2's version of it is to blame. :/

Apache 2's algorithm is different -- I think it is a coding error, not
intentional.  Apache 1 does what I describe above.  Apache 2 tries to
read with a 2 second timeout up to 30/2 times, and each read is 512
bytes max.  That means if the socket's receive buffer _already_ has
512*15 bytes in it, Apache 2 will terminate the lingering close
immediately, even if there is more data incoming from the client.
That seems very wrong: incoming data from the client after close is
what causes the transmitted data to be lost, which is why Apache 1
keeps trying until it sees a 2 second gap, which heuristically means
the client has stopped sending.

> Also, FWIW, I recall half-close causing many problems
> for Squid proxies for a while. It is probably fixed now.

I don't see how it could be a problem: server half-close with a
timeout (the 30 seconds timeout is still important) is invisible to
the client, except in terms of the timing.  The sequence of events the
client is able to observe is identical (unless the client is _so_
clever that is queries the socket to learn how much data has been
acknowledged by the server, but I'm sure Squid doesn't do that).

In effect, half-close with a 30 second timeout is equivalent, at the
socket interface level, to the network deciding to delay data flowing
from the client to the server for 30 seconds, so that the TCP RST
effect where response data disappears can't happen.  A real network
can cause a similar effect, so if half-close was causing Squid
problems, those problems would occasionally occur with real network
delays too.

-- Jamie
Received on Tuesday, 16 March 2004 01:00:39 UTC

This archive was generated by hypermail 2.3.1 : Tuesday, 1 March 2016 11:10:37 UTC