decred / dcrlnd

Decred Lightning Network Daemon ⚡️
MIT License
36 stars 24 forks source link

Dependency cycle causing shutdown deadlock #97

Open matheusd opened 4 years ago

matheusd commented 4 years ago

Sub-systems are shutdown concurrently: https://github.com/decred/dcrlnd/blob/29759788a0212a4777aac8a7bd0f4410a9a9b3a6/server.go#L1411

However there's interdependency between them. Specifically, chanRouter depends on the current cc.chainViewinstance but calls to GetBlock might block indefinitely until the chainView is called to stop.

One situation where this is manifested is on the garbage_collect_link_nodes itest where a channel is opened between Alice and Dave and immediately after that Bob and Carol are restarted. It might be the case that Bob (or Carol) is in the middle of a GetBlock() call and on the remotewallet implementation that cashes out to a loop that is only returned from once the wallet is told to shutdown. But that only happens later in the server shutdown routine, so the wallet shutdown deadlocks.

Currently this is solved in the itest with a dummy Sleep call but ideally this should be solved either by performing subsystem shutdown in parallel (so that the wallet shutting down bubbles up an error to chanRouter) or by requiring the chainView methods to accept a context parameter, so that the caller itself can cancel them.