Description

klever-go: Unbounded goroutine spawn on direct-message ingress enables peer-driven DoS

Summary

networkMessenger.directMessageHandler in network/p2p/libp2p/netMessenger.go spawns a fresh goroutine for every incoming direct message before the antiflood layer makes an admission decision. There is no semaphore, throttler, or bound on concurrent in-flight spawns.

A single connected libp2p peer can open a DirectSendID stream and send well-formed TopicMessage envelopes with varying sequence numbers. Each accepted direct message reaches directMessageHandler and triggers a fresh goroutine before processor.ProcessReceivedMessage runs. This allows unbounded goroutine growth and node availability degradation from one peer.

This remains present in the latest release v1.7.17: network/p2p/libp2p/netMessenger.go:1060 still spawns go func(msg p2p.MessageP2P) before processor.ProcessReceivedMessage. I also verified current develop commit 10bcfd50, where the same spawn remains at network/p2p/libp2p/netMessenger.go:1115.

This is distinct from GHSA-74m6-4hjp-7226 and GHSA-87m7-qffr-542v. Those advisories concern MultiDataInterceptor decompression/throttler behavior. This report concerns the libp2p direct-message ingress wrapper spawning an unbounded goroutine before processor-level antiflood/admission logic runs. A patch to Batch.Decompress or MultiDataInterceptor does not bound this direct-message goroutine spawn.

Details

The affected path is network/p2p/libp2p/netMessenger.go in directMessageHandler.

The direct-message path transforms and validates the message, looks up the topic processor, then immediately spawns a goroutine:

func (netMes *networkMessenger) directMessageHandler(message *pubsub.Message, fromConnectedPeer core.PeerID) error {
    var processor p2p.MessageProcessor

    topic := *message.Topic
    msg, err := netMes.transformAndCheckMessage(message, fromConnectedPeer, topic)
    if err != nil {
        return err
    }...
Show more

The processor-level antiflood decision happens inside ProcessReceivedMessage, after the goroutine, its stack, and the cloned message reference already exist. That means antiflood can bound processing rate, but not goroutine creation rate.

The existing goRoutinesThrottler with capacity broadcastGoRoutines = 1000 is wired into outgoing broadcast paths such as BroadcastOnChannelBlocking, not this incoming direct-message path.

The parallel pubsub ingress path in the same file handles a comparable inbound message surface synchronously:

err = handler.ProcessReceivedMessage(msg, fromConnectedPeer)

So the direct-message path is asymmetric: same transform/check function, same ProcessReceivedMessage callee, but direct-message ingress adds an unbounded goroutine spawn.

Reachability:

directSender.go registers DirectSendID as a libp2p stream protocol.

directStreamHandler reads framed pubsub.Message envelopes from the stream.

directStreamHandler forwards each message to networkMessenger.directMessageHandler.

Any connected peer can send well-formed envelopes to registered topics.

The seenMessages cache keys on From + Seqno; Seqno is attacker-controlled in the envelope, so incrementing it bypasses dedupe.

PoC

GitHub Private Vulnerability Reporting does not appear to allow file attachments in this form, so I am including the reproduction command and captured output inline. I can provide the full Go test file immediately if useful.

The PoC is a Go test file intended to be placed under network/p2p/libp2p/ in a klever-go checkout. It exercises the real network/p2p/libp2p package with NewMockMessenger.

Reproduction:

git clone https://github.com/klever-io/klever-go
cd klever-go
git checkout v1.7.16

# network/p2p/libp2p/

go test ./network/p2p/libp2p/ -run TestPoC_ -count=1 -v -timeout 60s

Captured output:

=== RUN   TestPoC_DirectMessageHandler_SpawnsGoroutinePerMessage
    baseline goroutines: 43
    peak goroutines after 500 direct messages: 543 (delta = 500)
    final goroutines after drain + GC: 43
POC_RESULT direct=spawn N=500 baseline=43 peak=543 delta=500 threshold=400 final=43
--- PASS

=== RUN   TestPoC_SynchronousHandler_NoGoroutineGrowth...
Show more

Reading:

500 direct messages with slow processors produced exactly 500 new goroutines.

The synchronous control path produced zero goroutine growth.

2000 messages, twice the outgoing broadcastGoRoutines = 1000 capacity, returned immediately, showing no ingress throttler blocks this path.

Impact

A single connected peer can sustain unbounded goroutine spawn growth on a klever-go node. Each spawned goroutine allocates its own stack, holds message references until the processor returns, and adds scheduler and GC pressure before antiflood admission can reject the message.

Under realistic attacker line rate and non-trivial processor latency, goroutine count can grow faster than the runtime drains it, degrading the node's ability to process legitimate traffic. This maps to the SECURITY.md High category: "Denial of Service affecting network availability."

All testing was local only. I did not contact Klever mainnet, public testnet, hosted RPCs, explorers, or third-party production infrastructure.

Suggested fixes:

Wire goRoutinesThrottler.CanProcess() or a dedicated ingress throttler before the go func() spawn in directMessageHandler.

Or remove the goroutine and call ProcessReceivedMessage synchronously, matching the existing pubsubCallback path.

Disclosure note: I originally sent this report to [email protected] on 2026-05-13. Since SECURITY.md lists GitHub Private Vulnerability Reporting as the recommended channel, I am resubmitting it here.

Mitigation

Aliases

References