nagayon-935 / mping / 28497976334

85%

Build Type

push

github

Committed by nagayon-935

Commit Message

fix: discard torn discover() rounds instead of reporting a false flap

CI was failing intermittently on TestEngine_OnFlap_NotCalledOnSameRoute
(internal/mtr). Root cause: discover()'s per-TTL goroutines each check
ctx.Err() independently before appending their result. When Stop()'s
context cancellation races a RediscoverEvery tick, goroutines that check
ctx.Err() before cancellation append their result while ones that check
after it do not, producing a torn result set for that round — e.g. the
TTL that actually reached the destination can be dropped while
unresponsive higher TTLs still get registered, making the hop count
jump (1 -> cfg.MaxHops) and RouteSnapshot/CheckFlap see the round as a
spurious path-length change.

discover() now checks ctx.Err() once after wg.Wait() and discards the
whole round (returns 0, its only otherwise-unused value) rather than
registering a partial view. runTarget's rediscover branch treats that
sentinel the same way it already treats ctx.Err() at the initial
discovery call site: skip the flap check and keep the previous
hopCount, letting the next loop iteration observe ctx.Done() normally.

Reproduced locally (~7% failure rate over 30 runs); 100/100 clean after
the fix, and 20x -race runs of the mtr package pass.

Coverage Stats

12 of 20 new or added lines in 1 file covered. (60.0%)

164 existing lines in 7 files now uncovered.

4612 of 5514 relevant lines covered (83.64%)

0.98 hits per line