W3C home > Mailing lists > Public > ietf-http-wg@w3.org > April to June 2014

Re: Connection fatal HTTP/2 stream errors

From: Tatsuhiro Tsujikawa <tatsuhiro.t@gmail.com>
Date: Wed, 18 Jun 2014 18:55:33 +0900
Message-ID: <CAPyZ6=Li-zaG+C2ihgJrMAqmuswqpwEJe1g6JN9KT8RjGuiasA@mail.gmail.com>
To: Greg Wilkins <gregw@intalio.com>
Cc: HTTP Working Group <ietf-http-wg@w3.org>
Hi,


On Wed, Jun 18, 2014 at 6:32 PM, Greg Wilkins <gregw@intalio.com> wrote:

>
>
> We are currently working on improving Jetty's handing of stream errors and
> I'm wondering what other server implementations are doing about stream
> errors.
>
> The issue is that there are many errors that can be detected during hpack
> decoding, or soon after that could be handled with a RST_STREAM.  However
> because of either the difficult and/or expense of needing to continue with
> decoding the headers (to keep the hpack table state correct), we are
> currently leaning towards treating all stream errors as fatal to the
> connection.
>
> Such errors that I have identified include are:
>
>    - headers too large
>    - upper case in header name
>    - non US-ASCII character in header name or value
>    - unknown scheme
>    - non token bad method (spaces etc)
>    - badly formatted authority and/or bad port number
>    - bad path (eg non hex % encoded or bad character encoded)
>    - bad status
>
> Note that some of these checks could be done after hpack decoding, however
> if we are storing fields in the Hpack table, we try to store the decoded
> version of them rather than just the string.
>
We are looking setting an error state and continuing decoding, but in the
> case of too large that is a bit counter productive.  So currently we are
> thinking that any of these kinds of errors are going to result in a GO_AWAY
> message as it is just too difficult to handle such errors deep inside hpack
> decoding and continue.
>
> But the down side of this is that you can easily send poison pills that
> will kill a connection, which is pain if you are aggregating
> connections.
>
> Anyway, just wondering what other server implementations are trying to do.
> If others are going to make efforts to keep connections open and just
> ignore the bad streams, then I guess we will follow peer pressure and do
> the same.
>
> cheers
>
>
> ​​In nghttp2, we don't make them error except for headers too large case.
This can be easily accomplished just to store byte string to header table,
which is perfectly compliant to HPACK specification.
For large headers, the only limitation in nghttp2 is we limit the one
header name and value  ​must be less than 8KiB buffers.  The 8KiB is just a
magic and we are planning to add API to tweak its size.  If this limit
breaches, we treat it as connection error.  We may be able to do better
though.  Other than that, total number of header pairs are unlimited in
nghttp2 since we just buffer one header name/value pair only and we pass it
to application layer one by one.  The application layer can send back 413
(if it is server) or RST_STREAM if it thinks total number of headers or
size are too much to handle.

​Best regards,
Tatsuhiro Tsujikawa​



> ​




>
>
>
>
>
>
>
>
> --
> Greg Wilkins <gregw@intalio.com>
> http://eclipse.org/jetty HTTP, SPDY, Websocket server and client that
> scales
> http://www.webtide.com  advice and support for jetty and cometd.
>
Received on Wednesday, 18 June 2014 09:56:20 UTC

This archive was generated by hypermail 2.4.0 : Friday, 17 January 2020 17:14:31 UTC