* Michael Tokarev <mjt@xxxxxxxxxx> wrote: > Today (Friday, the 13th) I had a very bad sequence of failures > with our servers leading to data loss and almost the whole day > of very hard work. And now I'm *really* interested where the > fault(s) is(are). > > What I have here is an AMD780G-based system (Asus M3A-H/HDMI > motherboard, latest BIOS) with AND Phenom 9750 CPU and 8Gig of > ECC memory. The system is built for KVM (kernel virtual machine) > work, and is running several guests, but I'm not sure anymore > that KVM is related to the problem at hand. > > The problem is that - it seems - timekeeping on this machine is > quite unreliable. > > It's Phenom, so TSC should be synced. And it is being choosen > at bootup as clocksource. But regardless of current_clocksource > (tsc), it constantly increases hpet min_delta_ns, like this: > > Mar 13 19:58:02 gate kernel: CE: hpet increasing min_delta_ns to 15000 nsec > Mar 13 19:59:16 gate kernel: CE: hpet increasing min_delta_ns to 22500 nsec > Mar 13 19:59:16 gate kernel: CE: hpet increasing min_delta_ns to 33750 nsec > Mar 13 19:59:16 gate kernel: CE: hpet increasing min_delta_ns to 50624 nsec > Mar 13 20:47:02 gate kernel: CE: hpet increasing min_delta_ns to 75936 nsec > Mar 13 20:48:17 gate kernel: CE: hpet increasing min_delta_ns to 113904 nsec > Mar 13 21:02:23 gate kernel: CE: hpet increasing min_delta_ns to 170856 nsec > Mar 13 21:05:27 gate kernel: CE: hpet increasing min_delta_ns to 256284 nsec > Mar 13 21:07:28 gate kernel: Clocksource tsc unstable (delta = 751920452 ns) > Mar 13 21:09:12 gate kernel: CE: hpet increasing min_delta_ns to 384426 nsec > > and finally, it declares that TSC is unstable (pre-last line) and > switches to the (unstable) HPET. > > HPET min_delta_ns will be increasing further and further, i've seen it > increased to 576638 and more. > > And no doubt the system is unstable with KVM like crazy, especially under > some, even light, load. > > Today I were copying some relatively large amount of data over network from > another to this machine (to the host itself, not to any virtual guest), and > had numerous guest and host stalls and lockups. At times, host sops doing > anything at all, all guests stalling too, load average jumps to 80 and more, > and nothing happens. I can do something over console still, like running > top/strace, but nothing interesting shows. I captured Sysrq+T of this situation > here: http://www.corpit.ru/mjt/host-high-la -- everything I was able to find > in kern.log. 403 > After some time, sometimes it's several seconds, sometimes it's up to 10 > minutes, the thing "unstucks" and continues working. Today it happened after > about 10 minutes. But after it continued, 2 of the KVM guests were eating > 100% CPU and did not respond at all. The Sysrq+T of this is available at > http://www.corpit.ru/mjt/guest-stuck -- two KVM guests were not responsible. 403 too. > It's even more - the system started showing sporadic, random I/O > errors unrelated to the disks - for example, one of software RAID5 > arrays started behaving really oddly, so that finally, after a > reboot, I had to re-create the array and some of the filesystems > on it (which I never saw in last ~10 years I'm using sofraid on > linus, on many different systems and disks and with various > failure cases). > > Now, I switched to acpi_pm clocksource. And also tried to disable > nested page tables with kvm (kvm_amd npt=0). With that, > everything is slow and sluggish, but I was finally able to copy > that data without errors, while the guests were running. > > It were about to stuck as before, but I noticed it switched to > hpet (see "tsc is unstable" above) and I forced it to use acpi_pm > instead, and it survived. > > So, to the hell out of it all, and ignoring the magical Friday the 13th -- > who's fault it is? > > o why it declares tsc is unstable while phenom supposed to keep it ok? the TSC can drift slowly between cores, and it might not be in sync at bootup time already. You check check the TSC from user-space (on any kernel) via time-warp-test: http://redhat.com/~mingo/time-warp-test/MINI-HOWTO > o why hpet is malfunctioning? That's a question for Thomas i guess. > o why the system time on this machine is dog slow without special > adjtimex adjustments, while it worked before (circa 2.6.26) and > windows works ok here? > > For reference: > > https://sourceforge.net/tracker/?func=detail&atid=893831&aid=2351676&group_id=180599 > -- kvm bug on sourceforge, without any visible interest in even looking at it > > http://www.google.com/search?q=CE%3A+hpet+increasing+min_delta_ns > -- numerous references to that "CE: hpet increasing min_delta_ns" on the 'net, > mostly for C2Ds, mentioning various lockup issues > > http://marc.info/?t=123246270000002&r=1&w=2 -- > "slow clock on AMD 740G chipset" -- it's about the clock issue, also without > any visible interest. > > What's the next thing to do here? I for one don't want to see > todays failures again, it was very, and I mean *very* difficult > day to restore the functionality of this system that (and it isn't > restored at full because of the slowness of its current state). it's not clear which kernel you tried - if you tried a recent enough one then i'd chalk this up as a yet-unfixed timekeeping problem - which probably had ripple effects on KVM and the rest of the system. What would be helpful is to debug the problem :-) First verify that basic timekeeping is OK: does 'time sleep 10' really take precisely 10 seconds? Does 'date' advance precisely one second per physical second? A generic hw/sw state output of: http://people.redhat.com/mingo/cfs-scheduler/tools/cfs-debug-info.sh would also help people taking a look at this problem. If the problem persists, there might be a chance to debug it without rebooting the system. Rebooting and trying out various patches wont really work well for a server i guess. Ingo -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
