W3C home > Mailing lists > Public > ietf-http-wg@w3.org > April to June 2013

Re: Design Issue: Merge RST_STREAM and GOAWAY into a single ERROR frame type

From: James M Snell <jasnell@gmail.com>
Date: Fri, 3 May 2013 14:55:58 -0700
Message-ID: <CABP7Rbfi0Gj5Q3ED7gROwhgSwJ2P10G=2hjtdcZ9=AnUNAKJxw@mail.gmail.com>
To: William Chan (陈智昌) <willchan@chromium.org>
Cc: Roberto Peon <grmocg@gmail.com>, Jeff Pinner <jpinner@twitter.com>, "ietf-http-wg@w3.org" <ietf-http-wg@w3.org>
I've got no strong opinion on it either way, we just need to make sure
it is communicated clearly. Holding the position that all session
errors are always terminal seems perfectly reasonable given that such
errors will almost always be the result of either an improperly
implemented, malfunctioning or malicious endpoint.

On Fri, May 3, 2013 at 2:45 PM, William Chan (陈智昌)
<willchan@chromium.org> wrote:
> The lack of non-terminal session errors is reflective of the historical SPDY
> developer bias towards hard failure modes, so we don't tolerate
> implementation errors that may affect interoperability/correctness later on.
> It's opinionated and definitely has its issues.
>
> For "Session frame too big", my inclination would be to terminate the
> session. Google Chrome is not tolerant of things it thinks servers should
> not be doing :) I think it's been good for the ecosystem overall thus far.
> Are there other errors you'd like to see be non-terminal?
>
>
> On Fri, May 3, 2013 at 6:34 PM, James M Snell <jasnell@gmail.com> wrote:
>>
>> Very well.  With that, however, we still have the outstanding ed note...
>> How do we want to report non-terminal session errors (e.g.  Session frame
>> too big) or do we treat all session errors as terminal using goaway?
>>
>> On May 3, 2013 2:25 PM, "Roberto Peon" <grmocg@gmail.com> wrote:
>>>
>>> I also find the current way more obvious-- in wireshark and similar
>>> traces, it is far easier to pick out the different opcode type (which is
>>> typically rendered as the textual name of the opcode) as opposed to the
>>> numeric value in some field.
>>>
>>> -=R
>>>
>>>
>>> On Fri, May 3, 2013 at 2:12 PM, William Chan (陈智昌)
>>> <willchan@chromium.org> wrote:
>>>>
>>>> Sorry, my implication is that I don't see any objective determination of
>>>> what's simpler here, just subjective views of which many people can have an
>>>> opinion. But if there's consensus on doing this, then by all means, let's do
>>>> it. I for one disagree and find the current way simpler :)
>>>>
>>>>
>>>> On Fri, May 3, 2013 at 6:08 PM, James M Snell <jasnell@gmail.com> wrote:
>>>>>
>>>>> There is no bikeshedding going on at all. I made the motivation for
>>>>> this clear up front: it's a simplification that addresses three
>>>>> specific items. Note: there is an existing editorial note in the
>>>>> existing draft that calls out the fact that we have no non-terminal
>>>>> method of communicating non-stream related errors. If we can address
>>>>> that item while also simplifying things a bit, then fantastic.
>>>>>
>>>>> On Fri, May 3, 2013 at 2:01 PM, William Chan (陈智昌)
>>>>> <willchan@chromium.org> wrote:
>>>>> > This is a thread ripe for bikeshedding. Is there any major issue
>>>>> > worth
>>>>> > solving?
>>>>> >
>>>>> > If we're going to paint our bike sheds, my take is keep whatever
>>>>> > color the
>>>>> > bike shed already has unless it really offends a number of people.
>>>>> >
>>>>> >
>>>>> > On Fri, May 3, 2013 at 5:53 PM, James M Snell <jasnell@gmail.com>
>>>>> > wrote:
>>>>> >>
>>>>> >> Speaking candidly, if we find ourselves requiring more than 8
>>>>> >> boolean
>>>>> >> flags on an error frame we should all just quit and go home.
>>>>> >>
>>>>> >> On Fri, May 3, 2013 at 1:34 PM, Jeff Pinner <jpinner@twitter.com>
>>>>> >> wrote:
>>>>> >> > IIRC, when this was brought up at the last F2F the rational for
>>>>> >> > NOT
>>>>> >> > doing
>>>>> >> > this was that frame types were cheaper than flags (256 frame
>>>>> >> > types, 8
>>>>> >> > flags).
>>>>> >> >
>>>>> >> > That being said I think we should consider combining them :)
>>>>> >> >
>>>>> >> >
>>>>> >> > On Fri, May 3, 2013 at 1:04 PM, James M Snell <jasnell@gmail.com>
>>>>> >> > wrote:
>>>>> >> >>
>>>>> >> >> As a simplification, I'd like to suggest that we merge the
>>>>> >> >> RST_STREAM
>>>>> >> >> and GOAWAY frames into a single ERROR frame with the following
>>>>> >> >> definition:
>>>>> >> >>
>>>>> >> >>  0                   1                   2                   3
>>>>> >> >>  0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
>>>>> >> >> +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
>>>>> >> >> |                      Error Code (32)                          |
>>>>> >> >> +---------------------------------------------------------------+
>>>>> >> >> |X|                  Last-Stream-ID (31)                        |
>>>>> >> >> +-+-------------------------------------------------------------+
>>>>> >> >>
>>>>> >> >> (note that this flips the field order from the GOAWAY frame)
>>>>> >> >>
>>>>> >> >> A frame-specific GOAWAY flag bit (0x2) would be defined for the
>>>>> >> >> frame,
>>>>> >> >> and the Last-Stream-ID field would only be included in the frame
>>>>> >> >> data
>>>>> >> >> if this flag was set.
>>>>> >> >>
>>>>> >> >> This does a couple of things for us:
>>>>> >> >>
>>>>> >> >> 1. It simplifies the error handling and reduces the number of
>>>>> >> >> core
>>>>> >> >> frame
>>>>> >> >> types.
>>>>> >> >> 2. It allows us to terminate a stream and terminate the session
>>>>> >> >> in a
>>>>> >> >> single frame if necessary
>>>>> >> >> 3. It gives us a way of reporting non-terminal session errors
>>>>> >> >> (currently RST_STREAM is forbidden to use stream id #0 and GOAWAY
>>>>> >> >> is
>>>>> >> >> always terminal).
>>>>> >> >>
>>>>> >> >
>>>>> >>
>>>>> >
>>>>
>>>>
>>>
>
Received on Friday, 3 May 2013 21:56:44 UTC

This archive was generated by hypermail 2.3.1 : Tuesday, 1 March 2016 11:11:12 UTC