HeaderPropagation: reset AsyncLocal per request

[alefranz]

As Kestrel can bleed the AsyncLocal across requests,
see #13991.

Addresses #15384

/cc @benaadams @davidfowl

[benaadams]

Trouble with this is if there is an httprequest that happens with lag on a previous request due to fire and forgot (which would be likely problematic anyway) it will pick up the changed dictionary from the async local.

Either would need to make the Invoke method async which would detach the asynclocal; or make the base request method async to detach all asynclocals #14146

[alefranz]

Thanks Ben for the quick feedback.
I'll wait for the outcome of #14146.

[davidfowl]

I would also make this method async

[alefranz]

Changed approach and made the middleware async as suggested.
Not sure if it makes sense to go ahead with this change or wait for #14146 ?

[jkotalik]

@davidfowl didn't you fix this in 5.0 in Kestrel?

[halter73]

Why not just take #14027 instead?

[benaadams]

Why not just take #14027 instead?

Or #14146 which just makes ProcessRequest async so the changes are reset on each request

[alefranz]

would it be possible to include this in 3.1 given it is a bugfix, and eventually remove it for 5.0 once one of the global fixes lands?

[analogrelay]

We need to land on which of these we take, this PR or #14146 . Consensus in triage seems to lean towards #14146 .

As to backporting to 3.1, that will be open as an option once we merge to master. It's too late for 3.1.0 and this may be risky for a patch release, so we'll need to assess that separately.

[analogrelay]

We're going to close this in lieu of #14146 . If that doesn't land, we can revisit.