Re: HTTP/3 server push: handling of lost PUSH_PROMISE

[Kazuho Oku <[e-mail]>]

Tatsuhiro, thank you for raising the issue.

I actually wonder if the example being provided is a tip of an iceberg - is
the problem related to FIN at all?

Let's consider the following pattern:

* client sends a request
* server starts sending response, alongside a PUSH_PROMISE
* in addition, server initiates a push stream and starts sending data
* server-sent packets carrying the PUSH_PROMISE frame are lost
* client decides to cancel the request and sends RESET_STREAM & STOP_SENDING
* server continues sending the contents of the push stream

I could well be missing some aspects of push, but to me, the problem looks
like the lack of delivery guarantee of PUSH_PROMISE frames.


2021年7月27日(火) 17:56 Tatsuhiro Tsujikawa <[e-mail]>:

> Hi,
>
> On Tue, Jul 27, 2021 at 3:31 PM Martin Thomson <[e-mail]> wrote:
>
>> This is a case where QUIC processing something doesn't imply HTTP/3
>> processing something.  QUIC read the data and "processed" it.  HTTP/3
>> decided not to handle it deliberately, and the PUSH_PROMISE fell between
>> the cracks.
>>
>> One "solution" is to insist that endpoints attempt to process streams for
>> this stuff if they decide to discard responses.  This is like how in HTTP/2
>> you still have to update the HPACK table after resetting a stream.  It's
>> awkward, but I think that it would work if data is available.  I don't
>> think that there is any case in which data is unavailable to the client but
>> the server doesn't receive STOP_SENDING.
>>
>>
> It is indeed awkward that QUIC stack has to pass stream data all the way
> to the end of the stream (or sees RESET_STREAM or closed) to HTTP/3 client
> even after it requested stopping reading.  And the HTTP/3 client has to
> process it just for PUSH_PROMISE.  Does any implementation do this?
>
> Sending STOP_SENDING alone is not enough.  As I wrote in the previous
> post, by the time STOP_SENDING is received by the HTTP/3 server, it has
> finished processing the pushed stream and forgets it.  I imagine that if
> QUIC stack provides BSD socket-like interface and HTTP/3 server writes all
> data and can clear the memory without waiting for an acknowledgement.
>
> Best regards,
> Tatsuhiro Tsujikawa
>
>
>> On Tue, Jul 27, 2021, at 13:01, Tatsuhiro Tsujikawa wrote:
>> > Hi,
>> >
>> > It looks like in certain conditions, client is unable to process pushed
>> > stream and leaves it in unprocessable state indefinitely.
>> >
>> > Consider that client opens bidi stream and server sends PUSH_PROMISE
>> > and completes the response body which is very short (just a single
>> packet
>> > or two).  For some reason, client has decided to stop reading
>> > response, but FIN is seen (data recvd state), and does not send
>> > STOP_SENDING because it is not required (RFC mentions that it has a
>> > little value to send STOP_SENDING in data recvd state and
>> > unnecessary).  Client discards all stream data without handing it over
>> > to application, so PUSH_PROMISE is not processed.  Client does not
>> > know the push ID.  Because STOP_SENDING is not sent, server has no
>> signal
>> > which indicates PUSH_PROMISE is not processed, and opens a pushed
>> > stream.  Client receives pushed stream, but unable to find the
>> > corresponding PUSH_PROMISE.  It holds pushed stream until it sees
>> > PUSH_PROMISE, but it never come.  This causes the pushed stream to be
>> held by
>> > client indefinitely.
>> >
>> > Even if client sends STOP_SENDING to the bidi stream, it might not
>> > work if server finishes sending stream data and a pushed stream and
>> > forgets them before receiving STOP_SENDING.
>> >
>> > Best regards,
>> > Tatsuhiro Tsujikawa
>>
>>

-- 
Kazuho Oku