dangernoodle-io / breadboard / 26352247157

100%

push

github

fix(bb_mdns): flush batch synchronously on overflow instead of dropping (#310)

Reproducer: esp32-wroom32 single-core board running SW mining at prio 20
starves Core 0; the IDF mDNS browse callback fills the pending batch
synchronously faster than the 50 ms flush timer can drain it.  Original
code dropped events with a cumulative counter that grew by 16 per cycle
(+16, +32, +49… as seen in the field).

New behaviour: when batch_append_locked finds the batch full it
synchronously calls batch_do_flush_locked (the same code the timer
callback uses), resets the batch, cancels the pending one-shot timer,
appends the new event, and re-arms the timer.  BB_MDNS_BATCH_MAX
becomes a max-items-per-queue-item knob, not a drop threshold.

The only legitimate drop signal is now when both the batch item slot and
the dispatcher queue are full simultaneously ("queue+batch full, event
dropped") — the real overload condition.

The flush timer callback is refactored to share batch_do_flush_locked,
eliminating duplicated swap-and-send logic.  The old cumulative drop
counter is retained for the queue-full path; the per-burst reset noted
in the brief is left for a follow-up — the counter now only increments
on the rare true-overload path, so cosmetic churn is low value.

Tests: two new test cases in test_bb_mdns_coalesce.c
- test_coalesce_overflow_flush_32_events_two_batches: drives 32 peers
  back-to-back without firing the flush timer; asserts exactly two
  batches of 16 enqueued and all 32 per-peer callbacks fired once.
- test_coalesce_queue_full_flush_drops_and_drain_recovers: holds the
  simulated dispatcher queue full; verifies flush is rejected, no
  callbacks fire, and drain restores normal operation.
- test_coalesce_queue_and_batch_full_drops_and_recovers: batch full
  + queue held full simultaneously; verifies BB_ERR_NO_SPACE returned,
  drop_count increments, and drain recovers.

Host shim extended with queue_depth_cap / queue_depth_hold / queue_drain
controls an... (continued)

1334 of 1334 branches covered (100.0%)

Branch coverage included in aggregate %.

2268 of 2268 relevant lines covered (100.0%)

912.36 hits per line