Add a dedicated method for disconnecting TLS connections

[julianbrost]

Properly closing a TLS connection involves sending some shutdown messages so that both ends know that the connection wasn't truncated maliciously. Exchanging those messages can stall for a long time if the underlying TCP connection is broken. The HTTP connection handling was missing any kind of timeout for the TLS shutdown so that dead connections could hang around for a long time.

This PR introduces two new methods on AsioTlsStream, namely ForceDisconnect() which just wraps the call for closing the TCP connection and GracefulShutdown() which performs the TLS shutdown with a timeout similar to it was done in JsonRpcConnection::Disconnect() before.

As the lambda passed to Timeout has to keep the connection object alive, AsioTlsStream is changed to inherit from SharedObject, adding reference counting to it directly. Previously, it was already created as Shared<AsioTlsStream> everywhere. Thus, a good part of the first commit is changing that type across multiple files.

fixes #9986

Edit (@Al2Klimov)

closes #10216
closes #10219
closes #10220

[julianbrost]

Open questions:

• Do we want to try to call ForceDisconnect() directly in case a connection is shut down due to a timeout like "no messages received" on JSON-RPC connections?

By now, I'm pretty sure the answer is yes.

Evidence: Take two connected Icinga 2 nodes and break individual connections by dropping the packets specific to that connection in a firewall. Both nodes will detect that no messages were received and reconnect, however, the old connection remains in an established state:

root@satellite-b-2:/# netstat -tpn | grep 172.18.0.33
tcp6       0      0 172.18.0.31:5665        172.18.0.33:49584       ESTABLISHED 60/icinga2          
root@satellite-b-2:/# iptables -A INPUT -p tcp -s 172.18.0.33 --sport 49584 -j DROP
root@satellite-b-2:/# netstat -tpn | grep 172.18.0.33
tcp6       0      0 172.18.0.31:5665        172.18.0.33:49098       ESTABLISHED 60/icinga2          
tcp6       0  82558 172.18.0.31:5665        172.18.0.33:49584       ESTABLISHED 60/icinga2          
root@satellite-b-2:/# iptables -A INPUT -p tcp -s 172.18.0.33 --sport 49098 -j DROP
root@satellite-b-2:/# netstat -tpn | grep 172.18.0.33
tcp6       0  94323 172.18.0.31:5665        172.18.0.33:49098       ESTABLISHED 60/icinga2          
tcp6       0  82558 172.18.0.31:5665        172.18.0.33:49584       ESTABLISHED 60/icinga2          
tcp6       0      0 172.18.0.31:5665        172.18.0.33:39946       ESTABLISHED 60/icinga2          

That implies that there's probably a resource leak in that scenario (until the kernel decides that the connection is actually dead and returns an error for the socket operations).

Unverified theory of what might happen: JsonRpcConnection::Disconnect() could block here:

icinga2/lib/remote/jsonrpcconnection.cpp

Line 226 in 9a8620d

m_WriterDone.Wait(yc);

Which waits for JsonRpcConnection::WriteOutgoingMessages() to complete which might hang indefinitely in these send operations:

icinga2/lib/remote/jsonrpcconnection.cpp

Lines 112 to 120 in 9a8620d

	for (auto& message : queue) {
	size_t bytesSent = JsonRpc::SendRawMessage(m_Stream, message, yc);
	if (m_Endpoint) {
	m_Endpoint->AddMessageSent(bytesSent);
	}
	}
	m_Stream->async_flush(yc);

[Al2Klimov]

Thing to consider

• JsonRpcConnection: Don't drop client from cache prematurely #10210 (comment)
  • TL;DR: just close dead connections, probably via destructor (shouldn't need much code)

[julianbrost]

I resolved conflicts and started implementing JsonRpcConnection::ForceDisconnect(), still far from finished, but if anyone is eager to have a look, feel free.

[julianbrost]

I resolved conflicts and started implementing JsonRpcConnection::ForceDisconnect(), still far from finished

While continuing that work, I figured that this might become a bigger rework of JsonRpcConnection than I anticipated. So I decided that this is better done in a separate PR that I'll create tomorrow.

This PR on it's own should already be enough of an improvement on its own, after all it even fixes a problem in the HTTP connection handling.

Open questions:

• I'm not yet sure about the last two commits (d6953cc, e90acc5): they basically replace two calls of async_shutdown() with calls to the new GracefulShutdown(). In both instances, those calls are already guarded by higher-level timeouts (apilistener.cpp, apilistener.cpp, ifwapichecktask.cpp), the idea for replacing them would be to have just a single call to async_shutdown() in our code base that's properly guarded.

In that regard, I figured that adding comments to these calls why they are fine is good enough, especially when compared that was needed just to add a redundant timeout and spawn a pointless coroutine (e90acc5).

I'm leaving the PR in a draft state for the moment because I still want to answer a few detail questions regarding the two new disconnect methods (I removed a cancel() because I didn't see a good reason for it to exist, I wonder if shutdown() is even necessary after a successful async_shutdown() and once I'm sure about all this, some doc comments explaining the exact behavior certainly won't hurt).

[julianbrost]

I removed a cancel() because I didn't see a good reason for it to exist

I still couldn't find a good reason to call lowest_layer().cancel() before next_layer().async_shutdown(), which is a bit of a problem if I want to add comments explaining what the code does. Like what would we want to cancel on the lowest, i.e. TCP layer? And if something was actually cancelled, how much sense would it make to continue using the TCP connection? Would it be possible to actually cancel something in the middle of writing a TLS record and what would continuing with a new TLS record during the shutdown do?

I wonder if shutdown() is even necessary after a successful async_shutdown()

That on the other hand turned out to be necessary: as confirmed using strace, without calling lowest_layer().shutdown(), there would otherwise be no shutdown() syscall.

[yhabteab]

Why not use lowest_layer().cancel() here as before? Shouldn't this call force async_shutdown to return? Right now, if you force disconnect here, which basically means you literally destroy the socket file descriptor, the call to shutdown() on the underlying TCP layer becomes pointless since the socket no longer exists at that point. I know we suppress the errors, but shutdown() would return a bad file descriptor error in this case, which I don't think is right.

[Al2Klimov]

So ec isn't needed anymore, right?

[Al2Klimov]

Suggested change

	void GracefulDisconnect(boost::asio::io_context::strand strand, boost::asio::yield_context yc);
	void GracefulDisconnect(boost::asio::io_context::strand& strand, boost::asio::yield_context yc);

[yhabteab]

The copy constructor of strand say this :):

Meaning, it's just like a shared pointer. So, your suggestion is not wrong, but neither is the current implementation.

[yhabteab]

Actually, using a reference here wouldn't work because that method spawns a coroutine, and there's nothing that keeps the m_IoStrand object from the Http or Rpc class alive till that coroutine finishes.