On Mon, 2009-11-16 at 16:07 +0100, Thomas Gleixner wrote: > On Sat, 14 Nov 2009, Wu Zhangjin wrote: > > > From: Wu Zhangjin <wuzhangjin@xxxxxxxxx> > > > > trace_clock_local() is based on the arch-specific sched_clock(), in X86, > > it is tsc(64bit) based, which can give very high precision(about 1ns > > with 1GHz). but in MIPS, the sched_clock() is jiffies based, which can > > give only 10ms precison with 1000 HZ. which is not enough for tracing, > > especially for Real Time system. > > > > so, we need to implement a MIPS specific sched_clock() to get higher > > precision. There is a tsc like clock counter register in MIPS, whose > > frequency is half of the processor, so, if the cpu frequency is 800MHz, > > the time precision reaches 2.5ns, which is very good for tracing, even > > for Real Time system. > > > > but 'Cause it is only 32bit long, which will rollover quickly, so, such > > a sched_clock() will bring with extra load, which is not good for the > > whole system. so, we only need to implement a arch-specific > > trace_clock_local() for tracing. as a preparation, we convert it as a > > weak function. > > Hmm, I'm not convinced that this is really a huge overhead. > > First of all the rollover happens once every 10 seconds on a 800MHz > machine. > > Secondly we have a lockless implementation of extending 32bit counters > to 63 bit which is used at least by ARM to provide a high resolution > sched_clock implementation. See include/linux/cnt32_63.h and the users > in arch/ > > But that's a problem which can be discussed seperately and does not > affect the rest of the tracing infrastructure. I really would > recommend that you implement a sched_clock for the r4k machines based > on cnt32_63 and measure the overhead. Having a fine granular > sched_clock in general is probably not a bad thing. > Thanks very much for your prompt, Will have a look at cnt32_to_63() in include/linux/cnt32_to_63.h later. Thanks & Regards, Wu Zhangjin
