Problem

GarnetAadAuthenticator.IsAuthorized() is on the per-command hot path — IsAuthenticated (which forwards to IsAuthorized()) is checked twice per command via AdminCommands.cs:32 and AdminCommands.cs:131.

In a cluster CPU profile under AAD authentication, DateTime.get_UtcNow shows up as the single hottest method at ~62.7% exclusive CPU.

Fix

Three pieces:

1. Garnet.common.CoarseTimeProvider — an instance class wrapping any TimeProvider. Exposes UtcNow (DateTime) and UtcNowTicks (long); both read a cached value via a pure memory load. The cache refresh is scheduled via TimeProvider.CreateTimer so a FakeTimeProvider in tests drives refreshes when its clock is advanced — production and tests therefore exercise the same hot path. CoarseTimeProvider.System is the process-wide singleton backed by TimeProvider.System (amortizes the single background Timer across all callers).

2. Garnet.common.CoarseDateTime — kept as a thin static facade over CoarseTimeProvider.System for back-compat / convenience.

3. GarnetAadAuthenticator:

   • Constructor takes an optional CoarseTimeProvider (defaults to CoarseTimeProvider.System).
   • AadAuthenticationSettings exposes a public TimeProvider parameter and owns the CoarseTimeProvider lifecycle (reuses .System singleton when no override → zero per-settings Timer cost; owns and disposes a per-settings instance otherwise).
   • Authenticate() precomputes the token validity window in ticks once (_validFromTicks / _validToTicks).
   • IsAuthorized() reads _coarseTime.UtcNowTicks and compares long against long, skipping DateTimes per-call Kind-bit masking.

Microbenchmark

benchmark/BDN.benchmark/Auth/IsAuthorizedBenchmark.cs on Windows 11 24H2, i7-12700K, .NET 10, x64, Server GC, DOTNET_TieredPGO=0:

Environment.TickCount64 is a real win over DateTime.UtcNow (~14×) and dramatically simpler, but on Windows it still reads from KUSER_SHARED_DATA via a method call and is ~10× slower than the L1-cached Volatile.Read of a static long. On Linux the gap compresses (both TickCount64 and DateTime.UtcNow go through vDSO clock_gettime) but CoarseDateTime still wins. For a per-command hot path checked twice per call, CoarseDateTime.UtcNowTicks (0.12 ns ≈ a single mov) is the right choice.

Correctness considerations

• CoarseTimeProvider granularity is 100ms; AAD token validity windows are measured in hours, so the slop is operationally invisible (smaller than typical inter-host clock skew). A token that just expired may be accepted for up to ~100ms past its true deadline. Under sustained ThreadPool starvation the staleness can exceed 100ms; still harmless for hour-long token windows.
• GarnetAadAuthenticator is constructed once per RespServerSession (per client connection) and IsAuthorized() / Authenticate() are only ever called on that session''s RESP processing thread, so _validFromTicks / _validToTicks need no cross-thread synchronization.
• Failed auth in the catch block resets _validFromTicks = long.MaxValue / _validToTicks = long.MinValue so a previously-cached authorized state cannot leak past a failed re-authentication.
• The process-wide CoarseTimeProvider.System instance is static readonly, so it survives for process lifetime; ~10 callbacks/sec is negligible ThreadPool load. Cached value writes/reads use Volatile.Write / Volatile.Read of a long field — atomic on 64-bit, no JIT reordering.

Testing

• dotnet build clean.
• dotnet test test/cluster/Garnet.test.cluster --filter "FullyQualifiedName~Aad":
  • Existing ClusterAadAuthTests (×2) — full server AUTH flow via Authenticate() → IsAuthorized() and subsequent commands re-checking via IsAuthenticated. Pass on net8.0 and net10.0.
  • New GarnetAadAuthenticatorTests (×6) — unit-level tests using FakeTimeProvider (Microsoft.Extensions.TimeProvider.Testing) plumbed through CoarseTimeProvider. Covers default fallback to .System, valid window, expiry after Advance(), staying valid across multiple Advance() calls, malformed token, and failed re-auth clearing previously authorized state. Because the authenticator always reads _coarseTime.UtcNowTicks (no test-only branch), advancing the FakeTimeProvider fires the same TimeProvider.CreateTimer refresh callback the production cache uses — tests exercise the production hot path.
• dotnet format clean on the changed files.