Limit usage of HTTP/2 connections

[sbordet]

Jetty version
9.4.x

Description
HTTP/2 connections have an implicit max usage count determined by the fact that stream ids can go up to 2^31-1 and cannot be negative.

As such, the HTTP/2 client connection pool should default to close connections as they reach that limit (or maybe just before that).

[lorban]

@sbordet I've submitted a draft PR that enforces such limitation in HttpClientTransportOverHTTP2. Did you have something more sophisticated in mind?

[sbordet]

@lorban streams are incremented by 2, so it should be something like (2^31-1)/2 as the max usage count.
And we have to make sure that we don't accidentally go off-by-one (i.e. use the stream one more time that would produce a stream id of 2^31).

2^31-1 = 2147483647

On the client, stream usage N produces stream id N*2-1, so we should be safe with N=1073741823.
However, we need to take into account HTTP upgrades which have an implicit stream id of 1 -- not sure how that is counted.

On the server, stream usage N produces stream id N*2, so we should be safe with N=1073741823.

I would probably be on the safe side and stay at N=1073741822 or perhaps even less.

On the client, it's easy to set this value on the Pool.

On the server, though, there is no Pool so we need to explore some other mechanism to enforce this.

There are 2 choices for the server:
. stop pushing without closing the connection.
. gracefully close the connection so the client is forced to open a new one.

I'd go with the simpler, which I feel is stop pushing -- after all it's unlikely that the server pushes 1 billion resources to the client through the same connection (although it's possible).

We need test cases for the server case (I think the client case is already covered in the Pool tests).

[github-actions]

This issue has been automatically marked as stale because it has been a
full year without activity. It will be closed if no further activity occurs.
Thank you for your contributions.

[sbordet]

@lorban status of this issue?

[lorban]

@sbordet enforcing (and testing) this on the server is far from trivial so the balance between the benefits vs the required effort has been negative so far.

[sbordet]

@lorban I don't think it's that difficult. Can you detail?
BTW, don't we do this already for HTTP/3 as well?

[github-actions]

This issue has been automatically marked as stale because it has been a
full year without activity. It will be closed if no further activity occurs.
Thank you for your contributions.