dragonxd serves RPC with 8 workers and a 16-deep queue. While a witness
rebuild holds cs_main, every request needing it waits, and a request the
client gave up on keeps its slot. The GUI kept polling and re-probing
until all 24 slots were its own abandoned requests (even stop was then
refused), treated the busy node as lost (tearing down the wallet state),
and retried connect at frame rate on some pages.
Busy gate (src/rpc/rpc_busy_gate.h), enforced in RPCClient::performCall
for the shared rpc_ / fast_rpc_:
- the first reply timeout or queue-full refusal marks the node busy;
from then on nothing needing cs_main is sent (calls fail at once with
the new RpcFailureKind::Deferred). stop, help and getrescaninfo are
lock-free in dragonxd and always go through
- one canary getinfo with no client timeout waits on its own client and
thread; its answer clears the gate and triggers an immediate refresh
(queue-full: retried after 2..30 s; no answer: back to normal handling)
- the timer dispatch skips Core/Tx/Addresses/Peers/chat/mining/opid and
connect attempts while busy
- the 3-strike teardown counts only failures that put the connection in
doubt, never busy ones; no balance scans follow a busy chain-info call
Reconnect storm: retries were armed with setTimer(Core, kCoreDefault - 1),
i.e. "4 s elapsed", which against Overview's 2 s interval meant every
frame. RefreshScheduler::deferFor() is an absolute hold, and
ConnectRetryBackoff picks the delay by failure kind (not listening
1/2/5 s, queue full 2..30 s, busy 30 s, else 4 s).
Measured against the fake daemon: abandoned requests during a rebuild
went from ~24 (every slot) to at most 2, with no teardown and recovery
when the rebuild ends; a saturated queue sees about one canary per
back-off step instead of ~190 requests/s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
DanSandClaude Opus 5.5
0f25c67600