// мой код пока примерно такой:

    class TSingleProcessorMode
    {
        KDPC DpcTraps[MAXIMUM_PROCESSORS];
        volatile LONG Stall;
        KIRQL SavedIrql;
        CCHAR CpuCount;
        KPRIORITY SavedPriority;
        static void DpcRoutine(KDPC *pDpc, void *pContext, void *pArg1, void *pArg2);
    public:
        void Initialize();
        void Enter();
        void Exit();
    };

    void TSingleProcessorMode::Initialize()
    {
        RtlZeroMemory(this, sizeof(TSingleProcessorMode));
        CpuCount = (CCHAR) GetNumberOfProcessors();
        if(CpuCount > 1)
        {
            for(CCHAR i = 0; i < CpuCount; i++)
            {
                KeInitializeDpc(&DpcTraps[i], DpcRoutine, this);
                KeSetImportanceDpc(&DpcTraps[i], LowImportance);
                KeSetTargetProcessorDpc(&DpcTraps[i], i);
            }
        }
    }

    void TSingleProcessorMode::DpcRoutine(KDPC *pDpc, void *pContext, void *pArg1, void *pArg2)
    {
        TSingleProcessorMode *pThis = (TSingleProcessorMode *) pContext;

        InterlockedDecrement(&pThis->Stall);
        do
            __Crt::ProcessorPause();
        while(pThis->Stall > 0);

        if(pThis->Stall >= 0)
        {
            KIRQL Irql;
            KeRaiseIrql(HIGH_LEVEL, &Irql);
            DbgPrint("SingleProcessorMode: cpu %d stalled\n", KeGetCurrentProcessorNumber());
            do
                __Crt::ProcessorPause();
            while(pThis->Stall >= 0);
            KeLowerIrql(Irql);
        }
    }

    void TSingleProcessorMode::Enter()
    {
        DbgPrint("SingleProcessorMode: entering\n");
        if(CpuCount > 1)
        {
            SavedPriority = KeSetPriorityThread(KeGetCurrentThread(), HIGH_PRIORITY);
            KAFFINITY ActiveProcessors = KeQueryActiveProcessors();
            KeRaiseIrql(DISPATCH_LEVEL, &SavedIrql);
            CCHAR CurrentProcessor = (CCHAR) KeGetCurrentProcessorNumber();
            Stall = 1;
            for(CCHAR i = CpuCount - 1; i >= 0; i--)
                if(i != CurrentProcessor && (ActiveProcessors & (1u << i)) != 0)
                {
                    InterlockedIncrement(&Stall);
                    KeInsertQueueDpc(&DpcTraps[i], 0, 0);
                }
            KeLowerIrql(SavedIrql);
            InterlockedDecrement(&Stall);
            __Crt::ProcessorPause();
            KeRaiseIrql(HIGH_LEVEL, &SavedIrql);
            if(Stall > 0)
            {
                DbgPrint("SingleProcessorMode: wait for %d cpu\n", Stall);
                KeLowerIrql(SavedIrql);
                do
                    __Crt::ProcessorPause();
                while(Stall > 0);
                KeRaiseIrql(HIGH_LEVEL, &SavedIrql);
            }
        }
        else
            KeRaiseIrql(HIGH_LEVEL, &SavedIrql);
    }

    void TSingleProcessorMode::Exit()
    {
        if(CpuCount > 1)
        {
            InterlockedExchange(&Stall, -1);
            KeLowerIrql(SavedIrql);
            KeSetPriorityThread(KeGetCurrentThread(), SavedPriority);
        }
        else
            KeLowerIrql(SavedIrql);
        DbgPrint("SingleProcessorMode: exit\n");
    }