The v1.0.14 fix replaced one contended sync.RWMutex (RefreshCoordinator.
refreshMutex) with sync.Map. Production showed the same death-spiral
signature recurring ~2 hours later — same shape, different mutex:
65 goroutines stuck on a sync.(*RWMutex).Lock at one address, pod
pinned at 1000m CPU, identical Yaegi runCfg/reflect.Value.Call stack
pattern. The mutex was RefreshCoordinator.attemptsMutex.
Generalising: under Yaegi (interpreted Go for traefik plugins), any
per-request global mutex acquisition is a latent serialization point.
reflect.Value.Call dispatch on a held lock turns a microsecond
critical section into a multi-millisecond one, and on a GOMAXPROCS=1
pod the queue is unbounded.
This commit removes every per-request global mutex on the hot path:
1. RefreshCoordinator.attemptsMutex (sync.RWMutex)
sessionRefreshAttempts: map -> sync.Map.
refreshAttemptTracker: all fields atomic (int32, int64 UnixNano,
cooldownEndNano == 0 as the not-in-cooldown sentinel, replacing
the inCooldown bool).
isInCooldown / recordRefreshAttempt / recordRefreshSuccess /
recordRefreshFailure all become lock-free. Cooldown entry uses
CompareAndSwapInt64 so only one goroutine logs the transition.
2. RefreshCircuitBreaker.mutex (sync.RWMutex)
lastFailureTime / lastSuccessTime -> atomic.Int64 UnixNano.
state and failures already atomic.
AllowRequest / RecordSuccess / RecordFailure now pure atomic ops.
3. TraefikOidc.firstRequestMutex (sync.Mutex)
firstRequestReceived bool -> firstRequestStarted int32.
metadataRefreshStarted bool -> metadataRefreshStartedAtomic int32.
ServeHTTP bootstrap path uses CompareAndSwapInt32 — fires once,
zero steady-state cost. Previously the mutex was acquired on
every non-health request forever.
4. TraefikOidc.metadataRetryMutex (sync.Mutex)
lastMetadataRetryTime time.Time -> lastMetadataRetryNano int64.
The 30-second retry throttle is now a CAS on lastMetadataRetryNano.
cleanupStaleEntries iterates via sync.Map.Range; eviction is a
CompareAndDelete by pointer identity so a tracker freshly re-used by
a concurrent caller is not lost.
Empirical evidence (3 specialist-agent analysis of the v1.0.14 spike,
profiles in /tmp/traefik-spike-1779511683/):
* mutex profile: 97% delay in sync.(*Mutex).Unlock via
HTTPHandlerSwitcher -> accesslog -> metrics -> backoff.RetryNotify
* 65 stuck goroutines at one RWMutex address (0x40022eb648),
identical Yaegi CFG pointer, all on rc.attemptsMutex via
recordRefreshAttempt + isInCooldown
* traffic driver: long-lived in-cluster Go-http-client doing
~5.4 req/s POST embeddings via OIDC cookie session → same
sessionID → contention all funnels to one tracker entry
Yaegi support for sync/atomic confirmed at
github.com/traefik/yaegi@v0.16.1/stdlib/go1_22_sync_atomic.go:
AddInt32/Int64, LoadInt32/Int64, StoreInt32/Int64,
CompareAndSwapInt32/Int64 all exposed via reflect.ValueOf. Yaegi
dispatches each call through reflect.Value.Call to the COMPILED
atomic.* function, which executes a single hardware CAS/LOCK-XADD
instruction. Each atomic op still pays Yaegi dispatch cost but
cannot block — no queueing, no death spiral.
Trade-off acknowledged: v1.0.15 issues ~6-8 atomic/sync.Map ops per
leader-path request vs the 4 mutex ops of v1.0.14. Under low
contention this is a modest CPU bump. Under high contention it's
an unbounded → bounded transformation. Net win.
All tests pass with -race; golangci-lint clean.
Hot-path JWT verification rebuilt the public key on every call:
jwk -> ToRSAPublicKey -> x509.MarshalPKIXPublicKey -> pem.Encode
-> verifySignature -> pem.Decode -> x509.ParsePKIXPublicKey -> verify
Under yaegi this pinned a CPU when many concurrent dashboard panels
poll behind the middleware. The PEM round trip is pure waste.
* jwk.go: cache pre-parsed crypto.PublicKey per kid alongside the
raw JWKSet (parallel cache entry, same 1h TTL, invalidates together).
* jwt.go: split verifySignatureWithKey from verifySignature; existing
PEM-input entry point preserved for backchannel-logout callers.
* token_manager.go: VerifyJWTSignatureAndClaims now goes straight from
jwks cache to verifySignatureWithKey, no PEM round trip and no
per-request availableKids slice.
* universal_cache.go: token/JWK/session Get() takes RLock when the
entry is unexpired, so concurrent token verifications no longer
serialize on a single mutex. LRU semantics for general and metadata
caches are unchanged (tests cover the strict-LRU contract there).
* mocks: MockJWKCache, EnhancedMockJWKCache, mockJWKCacheForLogout,
staticJWKCache satisfy the extended interface.
lukaszraczylo