KeWaitForSingleObject延时，太离谱了吧

在系统线程里
VOID ThreadProc(PWDMDRV_DEVICE_EXTENSION pdx)
{
  LARGE_INTEGER liTimeout;
  KeQuerySystemTime(&liTimeout);
  while (TRUE)
  {
    liTimeout.QuadPart += (LONGLONG)1000000; // 100ms
    NTSTATUS status = KeWaitForSingleObject(&pdx->Event, Executive, KernelMode, FALSE, &liTimeout);
    LARGE_INTEGER liNow;
    KeQuerySystemTime(&liNow);
    // ...
  }
  // ...
}
输出的 liTimeout 和 liNow 如下：
0:  127155521257812500 2003-12-10 17:48:45.781 127155521257812500 2003-12-10 17:48:45.781
1:  127155521258812500 2003-12-10 17:48:45.881 127155521258906250 2003-12-10 17:48:45.890
2:  127155521259812500 2003-12-10 17:48:45.981 127155521259843750 2003-12-10 17:48:45.984
3:  127155521260812500 2003-12-10 17:48:46.081 127155521260781250 2003-12-10 17:48:46.078
4:  127155521261812500 2003-12-10 17:48:46.181 127155521261875000 2003-12-10 17:48:46.187
5:  127155521262812500 2003-12-10 17:48:46.281 127155521262812500 2003-12-10 17:48:46.281
6:  127155521263812500 2003-12-10 17:48:46.381 127155521263750000 2003-12-10 17:48:46.375
7:  127155521264812500 2003-12-10 17:48:46.481 127155521264843750 2003-12-10 17:48:46.484
8:  127155521265812500 2003-12-10 17:48:46.581 127155521265781250 2003-12-10 17:48:46.578
9:  127155521266812500 2003-12-10 17:48:46.681 127155521266875000 2003-12-10 17:48:46.687
左边是liTimeout，右边是liNow，运行环境：C1.7G，Win2000Server(SP4)，误差也太大了吧，KeDelayExecutionThread也是同样效果，问题到底出在哪呢，另外想实现精确的1ms延时有更好的方法吗？

又仔细看了一下MSDN，原来“System time is typically updated approximately every ten milliseconds.”，但如何得到精确的时间呢？KeQueryTickCount、KeQueryInterruptTime、KeQueryPerformanceCounter效果好象都不是很好，还有上边那个WaitFor能实现1ms延时吗？没有一个好的时间标尺很难衡量！

刚用KeQueryPerformanceCounter做了个实验，延时1ms，得到如下输出：
050: 127155598950666250 2003-12-10 19:58:15.066  0.002514 ms
051: 127155598950676250 2003-12-10 19:58:15.067  0.001397 ms
052: 127155598950686250 2003-12-10 19:58:15.068  0.001956 ms
053: 127155598950696250 2003-12-10 19:58:15.069  0.001676 ms
054: 127155598950706250 2003-12-10 19:58:15.070  0.001397 ms
055: 127155598950716250 2003-12-10 19:58:15.071  0.001676 ms
056: 127155598950726250 2003-12-10 19:58:15.072  0.001397 ms
057: 127155598950736250 2003-12-10 19:58:15.073  0.001676 ms
058: 127155598950746250 2003-12-10 19:58:15.074  0.001676 ms
059: 127155598950756250 2003-12-10 19:58:15.075  0.001397 ms
060: 127155598950766250 2003-12-10 19:58:15.076  0.001956 ms
061: 127155598950776250 2003-12-10 19:58:15.077  0.001676 ms
062: 127155598950786250 2003-12-10 19:58:15.078  0.966045 ms
063: 127155598950796250 2003-12-10 19:58:15.079  1.905549 ms
064: 127155598950806250 2003-12-10 19:58:15.080  2.929143 ms
065: 127155598950816250 2003-12-10 19:58:15.081  3.919772 ms
066: 127155598950826250 2003-12-10 19:58:15.082  4.872128 ms
067: 127155598950836250 2003-12-10 19:58:15.083  5.864712 ms
068: 127155598950846250 2003-12-10 19:58:15.084  0.001956 ms
069: 127155598950856250 2003-12-10 19:58:15.085  0.001397 ms
070: 127155598950866250 2003-12-10 19:58:15.086  0.001676 ms
071: 127155598950876250 2003-12-10 19:58:15.087  0.001397 ms
072: 127155598950886250 2003-12-10 19:58:15.088  0.001676 ms
073: 127155598950896250 2003-12-10 19:58:15.089  0.001676 ms
074: 127155598950906250 2003-12-10 19:58:15.090  0.001676 ms
075: 127155598950916250 2003-12-10 19:58:15.091  0.001676 ms
076: 127155598950926250 2003-12-10 19:58:15.092  0.001956 ms
077: 127155598950936250 2003-12-10 19:58:15.093  0.001676 ms
078: 127155598950946250 2003-12-10 19:58:15.094  0.953753 ms
079: 127155598950956250 2003-12-10 19:58:15.095  1.971200 ms
080: 127155598950966250 2003-12-10 19:58:15.096  2.909588 ms
081: 127155598950976250 2003-12-10 19:58:15.097  3.913626 ms
082: 127155598950986250 2003-12-10 19:58:15.098  4.875759 ms
最右边那列间隔显然是不能用的，头大！

NT内核本来就不是实时的，所以别太挑剔了。。。。。

看来还是KTIMER+KDPC效果好些
LARGE_INTEGER liDueTime;
liDueTime.QuadPart = (LONGLONG)-10000;
KeSetTimerEx(&pdx->Timer, liDueTime, (LONG)1, &pdx->Dpc);
基本上还是在2ms左右，间或也有超过4ms的，基本还可以接受，不明白别人us级的延时是如何做的

用queryperformancecounter和queryperformancefrequency可以计算出一毫秒需要的tick数，但不知道你怎么定时法，用timer不行吗？我没试过，你可以考虑一下啊，精度不够？

感谢关注！
我想KeQueryPerformanceCounter应该是可信的，精确延时可能不太现实，但ms级应该不算太苛刻吧，相当多的应用要求都比这高，看来单任务、硬件实现确实有其无法比拟的优越性！

您是说微秒级的延时？！