kernel: don't hold TaskSet.mu during most of Kernel.runCPUClockTicker()

The removed `tg.leader == nil` check doesn't actually affect the correctness of
the rest of the loop body.

PiperOrigin-RevId: 681163998

pkg/sentry/kernel/task_exit.go

@@ -737,13 +737,13 @@ func (t *Task) exitNotifyLocked(fromPtraceDetach bool) {
 			ns.deleteTask(t)
 		}
 		t.userCounters.decRLimitNProc()
-		t.tg.exitedCPUStats.Accumulate(t.CPUStats())
-		t.tg.ioUsage.Accumulate(t.ioUsage)
 		t.tg.signalHandlers.mu.Lock()
 		t.tg.tasks.Remove(t)
 		t.tg.tasksCount--
 		tc := t.tg.tasksCount
+		t.tg.exitedCPUStats.Accumulate(t.CPUStats())
 		t.tg.signalHandlers.mu.Unlock()
+		t.tg.ioUsage.Accumulate(t.ioUsage)
 		if tc == 1 && t != t.tg.leader {
 			// Our fromPtraceDetach doesn't matter here (in Linux terms, this
 			// is via a call to release_task()).

pkg/sentry/kernel/task_sched.go

@@ -204,7 +204,7 @@ func (tg *ThreadGroup) CPUStats() usage.CPUStats {
 }
 
 // Preconditions: Same as TaskGoroutineSchedInfo.userTicksAt, plus:
-//   - The TaskSet mutex must be locked.
+//   - Either the TaskSet mutex or the signal mutex must be locked.
 func (tg *ThreadGroup) cpuStatsAtLocked(now uint64) usage.CPUStats {
 	stats := tg.exitedCPUStats
 	// Account for live tasks.
@@ -375,12 +375,8 @@ func (k *Kernel) runCPUClockTicker() {
 				continue
 			}
 
-			k.tasks.mu.RLock()
-			if tg.leader == nil {
-				// No tasks have ever run in this thread group.
-				k.tasks.mu.RUnlock()
-				continue
-			}
+			sh := tg.signalLock()
+
 			// Accumulate thread group CPU stats, and randomly select running tasks
 			// using reservoir sampling to receive CPU timer signals.
 			var virtReceiver *Task
@@ -415,7 +411,6 @@ func (k *Kernel) runCPUClockTicker() {
 
 			// All of the following are standard (not real-time) signals, which are
 			// automatically deduplicated, so we ignore the number of expirations.
-			tg.signalHandlers.mu.Lock()
 			// It should only be possible for these timers to advance if we found
 			// at least one running task.
 			if virtReceiver != nil {
@@ -445,9 +440,8 @@ func (k *Kernel) runCPUClockTicker() {
 					profReceiver.sendSignalLocked(SignalInfoPriv(linux.SIGKILL), true)
 				}
 			}
-			tg.signalHandlers.mu.Unlock()
 
-			k.tasks.mu.RUnlock()
+			sh.mu.Unlock()
 		}
 
 		k.cpuClockMu.Unlock()

pkg/sentry/kernel/thread_group.go

@@ -204,7 +204,11 @@ type ThreadGroup struct {
 	nextTimerID linux.TimerID
 
 	// exitedCPUStats is the CPU usage for all exited tasks in the thread
-	// group. exitedCPUStats is protected by the TaskSet mutex.
+	// group. exitedCPUStats is protected by both the TaskSet mutex and the
+	// signal mutex. Mutating it requires that the TaskSet mutex is locked for
+	// writing *and* that the signal mutex is locked. Reading it requires
+	// locking the TaskSet mutex (for reading or writing) *or* locking the
+	// signal mutex.
 	exitedCPUStats usage.CPUStats
 
 	// childCPUStats is the CPU usage of all joined descendants of this thread