On 10.07.2013, at 00:26, Scott Wood wrote: > On 07/09/2013 05:00:26 PM, Alexander Graf wrote: >> On 09.07.2013, at 23:54, Scott Wood wrote: >> > On 07/09/2013 04:49:32 PM, Alexander Graf wrote: >> >> Not sure I understand. What the timing stats do is that they measure the >> >> time between [exit ... entry], right? We'd do the same thing, just all in >> >> C code. That means we would become slightly less accurate, but gain >> >> dynamic enabling of the traces and get rid of all the timing stat asm >> >> code. >> > >> > Compile-time enabling bothers me less than a loss of accuracy (not just a >> > small loss by moving into C code, but a potential for a large loss if we >> > overflow the buffer) >> Then don't overflow the buffer. Make it large enough. > > How large is that? Does the tool recognize and report when overflow happens? > > How much will the overhead of running some python script on the host, > consuming a large volume of data, affect the results? > >> IIRC ftrace improved recently to dynamically increase the buffer size too. >> Steven, do I remember correctly here? > > Yay more complexity. > > So now we get to worry about possible memory allocations happening when we > try to log something? Or if there is a way to do an "atomic" log, we're back > to the "buffer might be full" situation. > >> > and a dependency on a userspace tool >> We already have that for kvm_stat. It's a simple python script - and you >> surely have python on your rootfs, no? >> > (both in terms of the tool needing to be written, and in the hassle of >> > ensuring that it's present in the root filesystem of whatever system I'm >> > testing). And the whole mechanism will be more complicated. >> It'll also be more flexible at the same time. You could take the logs and >> actually check what's going on to debug issues that you're encountering for >> example. >> We could even go as far as sharing the same tool with other architectures, >> so that we only have to learn how to debug things once. > > Have you encountered an actual need for this flexibility, or is it > theoretical?
Yeah, first thing I did back then to actually debug kvm failures was to add trace points. > Is there common infrastructure for dealing with measuring intervals and > tracking statistics thereof, rather than just tracking points and letting > userspace connect the dots (though it could still do that as an option)? > Even if it must be done in userspace, it doesn't seem like something that > should be KVM-specific. Would you like to have different ways of measuring mm subsystem overhead? I don't :). The same goes for KVM really. If we could converge towards a single user space interface to get exit timings, it'd make debugging a lot easier. We already have this for the debugfs counters btw. And the timing framework does break kvm_stat today already, as it emits textual stats rather than numbers which all of the other debugfs stats do. But at least I can take the x86 kvm_stat tool and run it on ppc just fine to see exit stats. > >> > Lots of debug options are enabled at build time; why must this be >> > different? >> Because I think it's valuable as debug tool for cases where compile time >> switches are not the best way of debugging things. It's not a high profile >> thing to tackle for me tbh, but I don't really think working heavily on the >> timing stat thing is the correct path to walk along. > > Adding new exit types isn't "working heavily" on it. No, but the fact that the first patch is a fix to add exit stats for exits that we missed out before doesn't give me a lot of confidence that lots of people use timing stats. And I am always very weary of #ifdef'ed code, as it blows up the test matrix heavily. Alex _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
