try and lock session get/update

pkg/auth/oauth/client.go

@@ -9,6 +9,7 @@ import (
 	"fmt"
 	"log/slog"
 	"strings"
+	"sync"
 	"time"
 
 	"atcr.io/pkg/atproto"
@@ -146,6 +147,7 @@ type UISessionStore interface {
 type Refresher struct {
 	clientApp      *oauth.ClientApp
 	uiSessionStore UISessionStore // For invalidating UI sessions on OAuth failures
+	didLocks       sync.Map       // Per-DID mutexes to prevent concurrent DPoP nonce races
 }
 
 // NewRefresher creates a new session refresher
@@ -162,8 +164,44 @@ func (r *Refresher) SetUISessionStore(store UISessionStore) {
 
 // GetSession gets a fresh OAuth session for a DID
 // Loads session from database on every request (database is source of truth)
+// Uses per-DID locking to prevent concurrent requests from racing on DPoP nonce updates
+//
+// Why locking is critical:
+// During docker push, multiple layers upload concurrently. Each layer creates a new
+// ClientSession by loading from database. Without locking, this race condition occurs:
+//   1. Layer A loads session with stale DPoP nonce from DB
+//   2. Layer B loads session with same stale nonce (A hasn't updated DB yet)
+//   3. Layer A makes request → 401 "use_dpop_nonce" → gets fresh nonce → saves to DB
+//   4. Layer B makes request → 401 "use_dpop_nonce" (using stale nonce from step 2)
+//   5. DPoP nonce thrashing continues, eventually causing 500 errors
+//
+// With per-DID locking:
+//   1. Layer A acquires lock, loads session, handles nonce negotiation, saves, releases lock
+//   2. Layer B acquires lock AFTER A releases, loads fresh nonce from DB, succeeds
 func (r *Refresher) GetSession(ctx context.Context, did string) (*oauth.ClientSession, error) {
-	return r.resumeSession(ctx, did)
+	// Get or create a mutex for this DID to prevent concurrent session loads
+	// This prevents DPoP nonce race conditions when multiple layers upload simultaneously
+	mutexInterface, _ := r.didLocks.LoadOrStore(did, &sync.Mutex{})
+	mutex := mutexInterface.(*sync.Mutex)
+
+	// Serialize session loading per DID
+	mutex.Lock()
+	defer mutex.Unlock()
+
+	slog.Debug("Acquired session lock for DID",
+		"component", "oauth/refresher",
+		"did", did)
+
+	session, err := r.resumeSession(ctx, did)
+	if err != nil {
+		return nil, err
+	}
+
+	slog.Debug("Released session lock for DID",
+		"component", "oauth/refresher",
+		"did", did)
+
+	return session, nil
 }
 
 // resumeSession loads a session from storage
@@ -213,8 +251,8 @@ func (r *Refresher) resumeSession(ctx context.Context, did string) (*oauth.Clien
 	}
 
 	// Set up callback to persist token updates to SQLite
-	// This ensures that when indigo automatically refreshes tokens,
-	// the new tokens are saved to the database immediately
+	// This ensures that when indigo automatically refreshes tokens or updates DPoP nonces,
+	// the new state is saved to the database immediately
 	session.PersistSessionCallback = func(callbackCtx context.Context, updatedData *oauth.ClientSessionData) {
 		if err := r.clientApp.Store.SaveSession(callbackCtx, *updatedData); err != nil {
 			slog.Error("Failed to persist OAuth session update",
@@ -223,10 +261,14 @@ func (r *Refresher) resumeSession(ctx context.Context, did string) (*oauth.Clien
 				"sessionID", sessionID,
 				"error", err)
 		} else {
-			slog.Debug("Persisted OAuth token refresh to database",
+			// Log session updates (token refresh, DPoP nonce updates, etc.)
+			// Note: updatedData contains the full session state including DPoP nonce,
+			// but we don't log sensitive data like tokens or nonces themselves
+			slog.Debug("Persisted OAuth session update to database",
 				"component", "oauth/refresher",
 				"did", did,
-				"sessionID", sessionID)
+				"sessionID", sessionID,
+				"hint", "This includes token refresh and DPoP nonce updates")
 		}
 	}
 	return session, nil