On Tue, Sep 25, 2012 at 10:12 AM, Gilles Chanteperdrix <[email protected]> wrote: > On 09/24/2012 03:33 PM, Henri Roosen wrote: >> Hi, >> >> I'm currently doing some tests on an Intel ATOM D2700 and try to run >> Xenomai threads on different cores. Currently using kernel 3.4.6, >> x86-32 SMP, and Xenomai 2.6.1 SMP. >> >> It seems the kernel detects the ATOM's TSC as unusable in SMP >> configuration and switches to the jiffies clock. This because there is >> an offset between both TSC counters of both cores. It seems even >> though the CPU has constant_tsc and nonstop_tsc, there is no guarantee >> for the counters to have the same value. I guess just the offset stays >> the same. >> >> My question is what are the implications for Xenomai not being able to >> have a TSC in SMP configuration, as Xenomai heavily relies on the TSC? > > Xenomai should not take what Linux does into accout, and continue to use > the TSC, you should see if that is the case by running > cat /proc/xenomai/timer
Yes, Xenomai is still running on the TSC: root@target:~# cat /proc/xenomai/timer status=on:setup=32:clock=187739919672:timerdev=lapic:clockdev=tsc Basically Xenomai is running in this SMP configuration. The direct problem I am currently seeing is that our code, which has only been used on single-core up to now, uses the TSC as a fast high resolution 'wall time' (rt_timer_tsc()). This fails when running Xenomai threads on different cores and the TSC of the cores don't have the same values. I'm sure Linux has a good reason to switch away from the TSC in this case. So I have to rephrase my initial question to: are there any implications for Xenomai-SMP when the TSC counters of the cores running Xenomai threads are of different value? I'm not sure whether there is a way to set the TSC's to the same value? That would solve all my questions and concerns. Intel documentation states it is done when the cores receive RESET, so I'm not sure whether this is applicable to the ATOM architecture. > > Something else you should check is whether Linux stops using the TSC > without CONFIG_IPIPE, because the fact that Linux declares the TSC > unusable may be an unwanted consequence of the I-pipe patch, for > instance the changes done for dynamic tsc detection. This has nothing to do with the IPIPE; Linux detects the same "TSC warp between CPUs" and switches away from the TSC for vanilla kernels running on this board. > > -- > Gilles. _______________________________________________ Xenomai mailing list [email protected] http://www.xenomai.org/mailman/listinfo/xenomai
