Disable read from client when HttpTunnelConsumer for client is set up

[masaori335]

Fix #7603.

The read from the client is enabled to watch the client abort, but this makes ATS signals EOS event to the HttpSM and goes HttpSM::state_watch_for_client_abort() handler.

trafficserver/proxy/http/HttpSM.cc

Lines 899 to 900 in 7af385f

	// Enable further IO to watch for client aborts
	ua_entry->read_vio->reenable();

When HttpSM sets up a HttpTunnelConsumer for a client, we can disable read to avoid the issue because there is nothing to read from the client. The EOS is handled by write op and signaled to HttpTunnel correctly.

[shinrich]

Looks good.

[bryancall]

Is there a reason we don't handle the EOS appropriately? Do we also disable reads from the client in the HTTP/2 case? If so, then how does that work for bidirectional communication with HTTP/2?

[masaori335]

Is there a reason we don't handle the EOS appropriately?

The HttpSM is in the state_watch_for_client_abort state and it kills HttpTunnel on EOS if ua_txn->allow_half_open() returns false. Checking what HttpTunnel is doing might be another approach, but I'm not sure it's better than this.

trafficserver/proxy/http/HttpSM.cc

Lines 975 to 993 in 42db8bb

	case VC_EVENT_EOS: {
	// We got an early EOS.
	NetVConnection *netvc = ua_txn->get_netvc();
	if (ua_txn->allow_half_open()) {
	if (netvc) {
	netvc->do_io_shutdown(IO_SHUTDOWN_READ);
	}
	ua_entry->eos = true;
	} else {
	ua_txn->do_io_close();
	ua_buffer_reader = nullptr;
	vc_table.cleanup_entry(ua_entry);
	ua_entry = nullptr;
	tunnel.kill_tunnel();
	terminate_sm = true; // Just die already, the requester is gone
	set_ua_abort(HttpTransact::ABORTED, event);
	}
	break;
	}

Do we also disable reads from the client in the HTTP/2 case? If so, then how does that work for bidirectional communication with HTTP/2?

With HTTP/2, this change disables the read of Http2Stream (just resets Http2Stream::read_vio which is used for receiving a request from the client on the stream).
The read op of HTTP/2 connection is still available.

[masaori335]

We tried this patch (on top of 8.1.2) and observed

1. background fill is working well
2. fewer client aborts
3. +2K active client connections

The 3. might be a side effect of disabling read. I'll try another approach.

[masaori335]

It turns out #7641 is a better approach because of fewer side effects.