Re: h2 padding

On Sep 4, 2014, at 1:25 AM, Willy Tarreau <> wrote:

> On Wed, Sep 03, 2014 at 02:22:29PM -0500, Jason Greene wrote:
>> On Sep 3, 2014, at 2:00 PM, Brian Smith <> wrote:
>>> On Tue, Sep 2, 2014 at 11:34 PM, Poul-Henning Kamp <> wrote:
>>>> Brian Smith writes:
>>>>> Consider an implementation that sends every frame in its own TCP
>>>>> packet, perhaps with a 1 minute delay between frames. [...]
>>>> If this was a joke, you forgot the smiley.
>>>> If it wasn't, please explain why we should even think about entertaining
>>>> the convenience of such an implementation,
>>> Pretty sure I am being trolled here, but in case I'm not: It is common
>>> for "security people" to give an exaggerated example to make a
>>> vulnerability obvious, in order to save time debating things like "is
>>> a millisecond too small to matter?" You can replace "1 minute" with "1
>>> second" or virtual any other non-zero period of time and you still
>>> have the same problem. Similarly, the problem still holds even if
>>> every frame isn't in its own TCP packet, as long as any frame gets
>>> split according to some function of the length of the padding of a
>>> frame.
>> I guess I don?t see how this makes a difference? If an implementation has the
>> ability to fit a frame and its payload on one packet, doesn?t it have the
>> ability to fit two frames on the same packet? Further, there is really no
>> guarantee that an H2 frame will not be split in a way that defeats padding in
>> the first place.
> There are many unoptimized implementations of many protocols which do :
>    write(socket, frame, length)
> with TCP_NODELAY set, resulting in exactly one packet + PUSH flag sent for
> each frame.

Sure but my point was that an implementation that uses padding has to be aware of and have control of the transport layer to ensure that padded data (whether it be a field or another frame) is always included with the payload (e.g. writev, cork, etc). Even then its still not perfect since small MSS/MTU could fragment a payload from its padding.
> You can even see this with some HTTP servers sending headers in multiple
> packets. I think this is the case Brian cares about.

Right, and such servers are not likely to use padding correctly.

> Willy

Jason T. Greene
WildFly Lead / JBoss EAP Platform Architect
JBoss, a division of Red Hat

Received on Thursday, 4 September 2014 16:16:24 UTC