首页 > 解决方案 > nmi 内核调用紧密循环

问题描述

我有一个测试程序

void task()
{
   long sum=0;
   while(true) {
      for(int i=0; i<100000;++i)
         for(int j=0; j < 10000; ++j)
            sum += j
   }
}
int main()
{
   boost::thread t(task);
   t.join();
   return true;
}

我使用 perf record -t -e intel_pt// 来分析线程。我看到很多对nmi的调用,例如每隔几毫秒调用一次 2(?) 。有没有办法摆脱这种类型的中断?Centos 7,内核 3.10.0,kernel.watchdog = 0。

根据@Zulan 的评论进行编辑:gcc 版本:4.8.5。编译行:g++ -I//include/ -L//lib -lboost_system -lboost_thread test.cpp 处理器 Intel Platinum 8xxx/Gold 6xxx 系列。
perf 脚本输出片段:

       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff816ac7f5 nmi ([kernel.kallsyms]) => ffffffff816ad500 do_nmi ([kernel.kallsyms])
       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff816ad54f do_nmi ([kernel.kallsyms]) => ffffffff816ad5f3 do_nmi ([kernel.kallsyms])
       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff816ad5fa do_nmi ([kernel.kallsyms]) => ffffffff8103dc10 is_debug_stack ([kernel.kallsyms])
       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff8103dc62 is_debug_stack ([kernel.kallsyms]) => ffffffff816ad5ff do_nmi ([kernel.kallsyms])
       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff816ad60e do_nmi ([kernel.kallsyms]) => ffffffff816ad558 do_nmi ([kernel.kallsyms])
       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff816ad573 do_nmi ([kernel.kallsyms]) => ffffffff8113a3f0 rcu_nmi_enter ([kernel.kallsyms])
       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff8113a41e rcu_nmi_enter ([kernel.kallsyms]) => ffffffff810ba170 __smp_mb__before_atomic ([kernel.kallsyms])
       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff810ba17a __smp_mb__before_atomic ([kernel.kallsyms]) => ffffffff8113a423 rcu_nmi_enter ([kernel.kallsyms])
       a.out 25192 [009]  3976.829805:          1     branches:  ffffffff8113a427 rcu_nmi_enter ([kernel.kallsyms]) => ffffffff810ba180 __smp_mb__after_atomic ([kernel.kallsyms])
       ...
       a.out 25192 [009]  3976.829807:          1     branches:  ffffffff8113a4aa rcu_nmi_exit ([kernel.kallsyms]) => ffffffff8113a48a rcu_nmi_exit ([kernel.kallsyms])
       a.out 25192 [009]  3976.829807:          1     branches:  ffffffff8113a48c rcu_nmi_exit ([kernel.kallsyms]) => ffffffff816ad592 do_nmi ([kernel.kallsyms])
       a.out 25192 [009]  3976.829807:          1     branches:  ffffffff816ad5ed do_nmi ([kernel.kallsyms]) => ffffffff816ad760 do_nmi ([kernel.kallsyms])
       a.out 25192 [009]  3976.829807:          1     branches:  ffffffff816ad76e do_nmi ([kernel.kallsyms]) => ffffffff816ac7fa nmi ([kernel.kallsyms])
       a.out 25192 [009]  3976.829807:          1     branches:  ffffffff816ac804 nmi ([kernel.kallsyms]) => ffffffff816ac23c restore_args ([kernel.kallsyms])
       a.out 25192 [009]  3976.829808:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)

另一个:

       a.out 25192 [009]  3976.829808:          1     branches:  ffffffff816b6718 irq_work_interrupt ([kernel.kallsyms]) => ffffffff8102f620 smp_irq_work_interrupt ([kernel.kallsyms])
       a.out 25192 [009]  3976.829808:          1     branches:  ffffffff8102f629 smp_irq_work_interrupt ([kernel.kallsyms]) => ffffffff81090d20 irq_enter ([kernel.kallsyms])
       a.out 25192 [009]  3976.829808:          1     branches:  ffffffff81090d29 irq_enter ([kernel.kallsyms]) => ffffffff8113a360 rcu_irq_enter ([kernel.kallsyms])
       a.out 25192 [009]  3976.829808:          1     branches:  ffffffff8113a3a0 rcu_irq_enter ([kernel.kallsyms]) => ffffffff8113a3b8 rcu_irq_enter ([kernel.kallsyms])
       a.out 25192 [009]  3976.829808:          1     branches:  ffffffff8113a3c0 rcu_irq_enter ([kernel.kallsyms]) => ffffffff81137e10 rcu_eqs_exit_common.isra.31 ([kernel.kallsyms])
       ...
       a.out 25192 [009]  3976.829809:          1     branches:  ffffffff8113a27d rcu_irq_exit ([kernel.kallsyms]) => ffffffff8113a260 rcu_irq_exit ([kernel.kallsyms])
       a.out 25192 [009]  3976.829809:          1     branches:  ffffffff8113a26e rcu_irq_exit ([kernel.kallsyms]) => ffffffff81090e4e irq_exit ([kernel.kallsyms])
       a.out 25192 [009]  3976.829809:          1     branches:  ffffffff81090e50 irq_exit ([kernel.kallsyms]) => ffffffff8102f653 smp_irq_work_interrupt ([kernel.kallsyms])
       a.out 25192 [009]  3976.829809:          1     branches:  ffffffff8102f654 smp_irq_work_interrupt ([kernel.kallsyms]) => ffffffff816b671d irq_work_interrupt ([kernel.kallsyms])
       a.out 25192 [009]  3976.829809:          1     branches:  ffffffff816b671d irq_work_interrupt ([kernel.kallsyms]) => ffffffff816ac1ed ret_from_intr ([kernel.kallsyms])
       a.out 25192 [009]  3976.829809:          1     branches:  ffffffff816ac234 retint_swapgs ([kernel.kallsyms]) => ffffffff816ac23c restore_args ([kernel.kallsyms])
       a.out 25192 [009]  3976.829809:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)

perf 脚本输出中 irq_return() 的 grep'g 显示每 20 毫秒 3 个 irq。

           a.out 25192 [009]  3976.809174:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           405764 task (a.out)
       a.out 25192 [009]  3976.809178:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           405764 task (a.out)
       a.out 25192 [009]  3976.829808:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)
       a.out 25192 [009]  3976.829809:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)
       a.out 25192 [009]  3976.850317:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           405764 task (a.out)
       a.out 25192 [009]  3976.850393:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           405764 task (a.out)
       a.out 25192 [009]  3976.850395:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           405764 task (a.out)
       a.out 25192 [009]  3976.870945:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           405764 task (a.out)
       a.out 25192 [009]  3976.870984:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)
       a.out 25192 [009]  3976.870985:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)
       a.out 25192 [009]  3976.891571:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)
       a.out 25192 [009]  3976.891609:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)
       a.out 25192 [009]  3976.891611:          1     branches:  ffffffff816ac26c irq_return ([kernel.kallsyms]) =>           40576f task (a.out)

在 grub 中,isolcpus、nohz_full、rcu_nocbs 设置在特定的 cpus 上,并且 nmi_watchdog=0。使用taskset -c a.out 运行的测试程序。

标签: linux-kernelredhatperfnmi

解决方案


首先,如果你想跟踪内核,你会在跟踪中得到很多内核函数调用,包括但不限于与 NMI 相关的调用。

如果您想像上面发布的那样跟踪您的用户空间程序,请使用

perf record -t -e intel_pt//u ...

最好查阅文档https://elixir.bootlin.com/linux/latest/source/tools/perf/Documentation/intel-pt.txt


推荐阅读