Messages in this thread | | | Date | Thu, 04 Sep 2014 23:22:29 +0200 | From | Paolo Bonzini <> | Subject | Re: [PATCH] KVM: x86: fix kvmclock breakage from timers branch merge |
| |
Il 04/09/2014 22:58, Thomas Gleixner ha scritto: > This is simply wrong.
It is.
> Now I have no idea why you think it needs to add xtime_sec. If the > result is wrong, then we need to figure out which one of the supplied > values is wrong and not blindly add xtime_sec just because that makes > it magically correct. > > Can you please provide a proper background why you think that adding > xtime_sec is a good idea?
It's not a good idea indeed. I didn't fully digest the 3.16->3.17 timekeeping changes and messed up this patch.
However, there is a bug in the "base_mono + offs_boot" formula, given that:
- bisection leads to the merge commit of John's timers branch
- bisecting within John's timers branch, with a KVM commit on top to make the code much easier to trigger, leads to commit cbcf2dd3b3d4 (x86: kvm: Make kvm_get_time_and_clockread() nanoseconds based, 2014-07-16).
- I backported your patch to 3.16, using wall_to_monotonic + total_sleep_time + xtime_sec (wtm+xtime_sec as in pre-cbcf2dd3b3d4 code, total_sleep_time from 3.16 monotonic_to_bootbased) and it works
- In v2 of the patch I fixed the bug by changing the formula "base_mono + offs_boot" to "offs_boot - offs_real" (and then adding xtime_sec separately as in the 3.16 backport), but the two formulas "base_mono + offs_boot" and "offs_boot - offs_real + xtime_sec" ought to be identical.
I find "offs_boot - offs_real + xtime" more readable than the alternative "base_mono + offs_boot + xtime_nsec", so the fix doubles as a cleanup for me and I'm fine with it. But something must be wrong in the timekeeping code.
Paolo
| |