Skip to content

eMule eSE LiveTV v0.70b-eSE7.2.0 (tombstone + gossip + watchdog)

Choose a tag to compare

View all tags
@diad87 diad87 released this 18 May 10:14
· 204 commits to main since this release
v0.70b-eSE7.2.0
cb0c2e4

eMule eSE LiveTV v0.70b-eSE7.2.0 — 2026-05-18

Minor-version bump (7.1.x → 7.2.0) because this adds a real network
feature, not just a fix: peer-to-peer gossip of dead-stream
notifications plus a local watchdog that catches broadcaster
crashes (which by definition can't send the notification themselves).

Both emule.exe and ese-server.exe changed.

Why

After yesterday's full day of crash-driven ghost streams, the
constraint became clear: Kad caches outlive the broadcaster by
hours. Every node that learned about your stream keeps echoing it
back to fresh searchers for ~5 h after you die. Local-only fixes
(TTL, registry sweep, prune intervals) help the broadcaster's own
machine but not anyone else's. To kill a ghost across the network
you need active propagation of the death signal.

OP_LIVE_END was already a defined opcode (0xC8) and the broadcaster
already sent it from StopBroadcast to its direct viewers. What was
missing:

  1. Receivers didn't do anything useful with it — they just removed
    the sender from their peer list, but kept the streamKey in their
    Kad directory.
  2. The notification stopped at one hop — mesh peers, super-seeders,
    and anyone discovering via Kad never heard.
  3. If the broadcaster crashed instead of stopping cleanly,
    OP_LIVE_END was never sent at all. Most ghost streams come from
    this path, not from graceful stops.

Changes

1. Tombstone map per peer

CLiveStreamManager now keeps a CMap<CString, DWORD> m_streamTombstones
mapping hex streamKeys to expiration tick (default 30 min TTL).

Lookups:

  • IsStreamTombstoned(streamKey) — called by CLiveKadBridge::GetKnownStreams
    and OnKadSearchResult to discard cached Kad echoes that try to
    resurrect streams we already know are dead.

Writes:

  • OnStreamEnded(streamKey, reason, fromPeer) — single entry point
    used by both the OP_LIVE_END handler and the watchdog.
  • Pruned each Process() tick (every minute).

2. OP_LIVE_END handler now does the work

ListenSocket.cpp:OP_LIVE_END
now reads <streamKey 16><reason 1> from the packet (was: just
removing the sender), then calls OnStreamEnded which:

  1. Tombstones the streamKey for 30 min.
  2. Leaves the stream if we were viewing it.
  3. Gossips the END to our mesh peers (broadcastPeers ∪ viewPeers,
    excluding the peer that informed us). Each receiver tombstones
    first and ignores duplicates, so propagation is bounded — no
    storms.

3. Crash-detection watchdog

CLiveStreamManager::Process
now tracks m_dwLastLiveActivity — refreshed on every chunk receive
AND every bitmap heartbeat. If we are viewing and the clock goes
silent for >90 s, we fabricate a synthetic OP_LIVE_END(reason=error)
locally:

OnStreamEnded(deadKey, ESE_END_ERROR, /*fromPeer=*/ NULL);

Tombstones locally and gossips to the mesh. Broadcaster crashed
silently → 90 s later, every direct viewer notices → 90+ε seconds
later, every mesh peer of those viewers notices → fan-out, no
caching node ever sees a fresh lastSeen because nobody's still
publishing the stream.

The 90 s threshold is generous on purpose: heartbeats fire every 1 s
and chunks every 2-4 s, so 22-90 missed beats = unambiguously dead,
not a 4G blip.

4. StopBroadcast notifies wider

CLiveStreamManager::StopBroadcast
used to send OP_LIVE_END only to m_broadcastPeers (direct viewers
subscribed via OP_LIVE_SUBSCRIBE). Now it ALSO sends to m_viewPeers
(in case we were a relay) and to every key in m_peerCounters (every
mesh peer we've talked to). Covers super-seeders that were
redistributing to others.

Also pre-tombstones our own streamKey before gossiping so an echo
that bounces back through the mesh doesn't trigger another round.

5. Kad result filter on read

CLiveKadBridge::OnKadSearchResult
and GetKnownStreams consult
IsStreamTombstoned and drop matching entries. Stale Kad echoes from
other nodes can no longer resurrect a stream we've already buried.

What this does NOT fix

Cold Kad nodes — machines that learned about a stream from a
search result but never connected to the mesh — won't receive the
gossip. They keep echoing the dead stream to fresh searchers until
their own local TTL expires (~5 h with vanilla eMule's
KADEMLIAREPUBLISHTIMEK = 24h and indexer-side prune at
~5 h half-life).

For those nodes, only time plus everyone's local
IsStreamTombstoned check on incoming results eventually evicts
the entry. Anyone running 7.2.0 won't display it; anyone running
older clients will, until Kad TTL.

Listed for v7.2.x if it bites:

  • Publish a deliberate "deprecated" Kad entry (bitrate=0 sentinel)
    more aggressively to overwrite the active one in cold caches.
    Partially exists already in UnpublishStream — could be extended
    to also publish on tombstone-from-gossip.

Hashes

  • emule.exe → 47139CB568A325F96A40B81E26893DE5788A52D151BEAF059BCC444B6B196B7C
  • ese-server.exe → A90C54E3FB6BE0A64F8B80400EBE33688AF2D87E121B370A5C8D9059F0A8C48A

(ese-server.exe rebuilt as part of the emule.sln pre-build step.
No source-level changes from v7.1.9 in the Node side.)

Upgrading from v7.1.9

Hot-swap supported. Both binaries changed.

  1. Close emule (kill emule.exe, ese-server.exe, and any leftover
    ffmpeg.exe).
  2. Replace BOTH binaries from the assets (or re-extract the ZIP).
  3. Reopen emule.

If you were broadcasting when you stopped, expect a tombstone to be
sent to whoever was watching as part of StopBroadcast. They'll
hide your channel from their lists for 30 min — which is correct,
because the streamKey of your next broadcast will be different.

Wire compatibility

OP_LIVE_END (opcode 0xC8) was already eSE-only. Vanilla eMule
0.70b ignores it. Older eSE versions process it (remove peer) but
don't gossip; they're consumers, not propagators. v7.2.0+ peers
form the gossip-active subset of the mesh, which is enough — most
networks will converge to all-7.2.0 within days as people update.

Carried over

  • v7.1.9 — ghost-channel + stale-Kad-entry cleanup
  • v7.1.8 — ~CUpDownClient → OnPeerDisconnected (use-after-free)
  • v7.1.7 — /live SyntaxError fix (final) + Cache-Control
  • v7.1.6 — bitrate-driven ABR variant + muted autoplay
  • v7.1.5 — orphan ffmpeg reap
  • v7.1.4 — paste-link redirect
  • v7.1.3 — preflight badges SyntaxError
  • v7.1.2 — Kad publish TAG_SOURCEIP
  • v7.1.1 — IPv4 localhost

GPL-2.0-only.

Assets 5
Loading