Profiling in llinux perf showed a couple hotspots in tracing crate, which boiled down to a couple things:

1. error_span!() was used everywhere, which is very convenient - it makes all log messages contain useful metadata, e.g. peer's address would be printed in every log line printed from the peer's task. However this causes overhead even when there's no logging at all seemingly. So I changed them all to lower level and manually propagated needed context to info!, warn! and error! log messages.
2. the order of filters seems to matter for performance. The (99.9% of the time) useless HTTP log layer was causing tracing overhead simply because the order of filters was suboptimal.