This discussion has been locked.
You can no longer post new replies to this discussion. If you have a question you can start a new discussion

A low-overhead real time counter consistent cross processor cores

I am new to ARM. 

I am using Xilinx RFSOC which in particular comprises 4 A53 cores (Armv8-Arunning Linux in SMP configuration.

I need a real-time clock with 1 us (or better) resolution that can be accessed from user-level code (running in AArch64 mode) and yield consistent readings across all cores.

I would like to be able to read the clock from application (EL0) code with minimal overhead, which implies that such reads should not be trapped by the O/S.

Preferably the counter should be readable via an atomic 64-bit read.

Browsing the various documents I found on the net, it seems that the ARM system counter provides the necessary functionality. The question is how to read it from user code. I found the following registers:

CNTCV, (supports reads via CNTReadBase)

CNTVCT, (accessible via 32-bit reads?)

CNTPCT_EL0 (probably trapped by the O/S?)

I have the following questions:

a) Does any of the above counters comply with my requirements (i.e. it is directly readable from EL0 code and provides consistent time stamps across the 4 A53 cores)? If not, is there another counter that I could use?

b) How exactly do I read the counter from user-level C code running under Linux?

Thanks,

  • It is difficult to give a direct answer. It depends on

      - Linux runs in non-secure EL0 as guest app

      - Linux runs in non-secure EL1 as guest OS

      - Linux runs in non-secure EL0 as host app

      - Linux run in non-secure EL2 as host OS.

    From Armv8-A architecture perspective, it is possible to read CNTPCT_EL0 in NS EL0. But in software practice, you have to consider the Linux virtualization is supported or not. CNTPCT_EL0 access will be trapped into OS or not.

    The last but not the least, it is not a good idea to read the counter in NS-EL0 to get the "real time" result even if you can access the counter from NS-EL0.
    All the factors such as Linux scheduler / cache miss / memory stall / interrupts will make you get the different test results for the same test case.

  • AFAIK, CNTPCT_EL0 refers to the local hardware timer and its frequency should be 1-50Mhz. So the time resolution is enough for you.

     

    "The CNTPCT_EL0 register reports the current count value. CNTKCTL_EL1 controls whether EL0 can access the system timer." Probably Yes, but for more detail, you should check the manual.

     

    Yes, I'm sure the hardware local timer has a consistent frequency across all the core, as this is the prerequisite to feasible the OS's scheduling.

    If is accessible from EL0, you can just use assembly in C code to read the register like 

    asm volatile ("mrs %0, cntp_tval_el0":"=r" (value));

     btw, I'm not sure you should read the value of cntp_tval or cntpct, please see the manual for their meaning.