Hello Thiago, On Fri, Feb 22, 2019 at 07:57:52PM -0300, Thiago Jung Bauermann wrote: > When testing DLPAR CPU add/remove on a system under stress, > pseries_cpu_die() doesn't wait long enough for a CPU to die: > > [ 446.983944] cpu 148 (hwid 148) Ready to die... > [ 446.984062] cpu 149 (hwid 149) Ready to die... > [ 446.993518] cpu 150 (hwid 150) Ready to die... > [ 446.993543] Querying DEAD? cpu 150 (150) shows 2 > [ 446.994098] cpu 151 (hwid 151) Ready to die... > [ 447.133726] cpu 136 (hwid 136) Ready to die... > [ 447.403532] cpu 137 (hwid 137) Ready to die... > [ 447.403772] cpu 138 (hwid 138) Ready to die... > [ 447.403839] cpu 139 (hwid 139) Ready to die... > [ 447.403887] cpu 140 (hwid 140) Ready to die... > [ 447.403937] cpu 141 (hwid 141) Ready to die... > [ 447.403979] cpu 142 (hwid 142) Ready to die... > [ 447.404038] cpu 143 (hwid 143) Ready to die... > [ 447.513546] cpu 128 (hwid 128) Ready to die... > [ 447.693533] cpu 129 (hwid 129) Ready to die... > [ 447.693999] cpu 130 (hwid 130) Ready to die... > [ 447.703530] cpu 131 (hwid 131) Ready to die... > [ 447.704087] Querying DEAD? cpu 132 (132) shows 2 > [ 447.704102] cpu 132 (hwid 132) Ready to die... > [ 447.713534] cpu 133 (hwid 133) Ready to die... > [ 447.714064] Querying DEAD? cpu 134 (134) shows 2 > > This is a race between one CPU stopping and another one calling > pseries_cpu_die() to wait for it to stop. That function does a short busy > loop calling RTAS query-cpu-stopped-state on the stopping CPU to verify > that it is stopped, but I think there's a lot for the stopping CPU to do > which may take longer than this loop allows. > > As can be seen in the dmesg right before or after the "Querying DEAD?" > messages, if pseries_cpu_die() waited a little longer it would have seen > the CPU in the stopped state.
> > I see two cases that can be causing this race: > > 1. It's possible that CPU 134 was inactive at the time it was unplugged. In > that case, dlpar_offline_cpu() calls H_PROD on that CPU and immediately > calls pseries_cpu_die(). Meanwhile, the prodded CPU activates and start > the process of stopping itself. It's possible that the busy loop is not > long enough to allow for the CPU to wake up and complete the stopping > process. The problem is a bit more severe since, after printing "Querying DEAD?" for CPU X, this CPU can prod another offline CPU Y on the same core which, on waking up, will call rtas_stop_self. Thus we can have two concurrent calls to rtas-stop-self, which is prohibited by the PAPR. > > 2. If CPU 134 was online at the time it was unplugged, it would have gone > through the new CPU hotplug state machine in kernel/cpu.c that was > introduced in v4.6 to get itself stopped. It's possible that the busy > loop in pseries_cpu_die() was long enough for the older hotplug code but > not for the new hotplug state machine. I haven't been able to observe the "Querying DEAD?" messages for the online CPU which was being offlined and dlpar'ed out. > > I don't know if this race condition has any ill effects, but we can make > the race a lot more even if we only start querying if the CPU is stopped > when the stopping CPU is close to call rtas_stop_self(). > > Since pseries_mach_cpu_die() sets the CPU current state to offline almost > immediately before calling rtas_stop_self(), we use that as a signal that > it is either already stopped or very close to that point, and we can start > the busy loop. > > As suggested by Michael Ellerman, this patch also changes the busy loop to > wait for a fixed amount of wall time. > > Signed-off-by: Thiago Jung Bauermann <bauer...@linux.ibm.com> > --- > arch/powerpc/platforms/pseries/hotplug-cpu.c | 10 +++++++++- > 1 file changed, 9 insertions(+), 1 deletion(-) > > I tried to estimate good amounts for the timeout and loop delays, but > I'm not sure how reasonable my numbers are. The busy loops will wait for > 100 µs between each try, and spin_event_timeout() will timeout after > 100 ms. I'll be happy to change these values if you have better > suggestions. Based on the measurements that I did on a POWER9 system, in successful cases of smp_query_cpu_stopped(cpu) returning affirmative, the maximum time spent inside the loop was was 10ms. > Gautham was able to test this patch and it solved the race condition. > > v1 was a cruder patch which just increased the number of loops: > https://lists.ozlabs.org/pipermail/linuxppc-dev/2017-February/153734.html > > v1 also mentioned a kernel crash but Gautham narrowed it down to a bug > in RTAS, which is in the process of being fixed. > > diff --git a/arch/powerpc/platforms/pseries/hotplug-cpu.c > b/arch/powerpc/platforms/pseries/hotplug-cpu.c > index 97feb6e79f1a..424146cc752e 100644 > --- a/arch/powerpc/platforms/pseries/hotplug-cpu.c > +++ b/arch/powerpc/platforms/pseries/hotplug-cpu.c > @@ -214,13 +214,21 @@ static void pseries_cpu_die(unsigned int cpu) > msleep(1); > } > } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) { > + /* > + * If the current state is not offline yet, it means that the > + * dying CPU (which is in pseries_mach_cpu_die) didn't have a > + * chance to call rtas_stop_self yet and therefore it's too > + * early to query if the CPU is stopped. > + */ > + spin_event_timeout(get_cpu_current_state(cpu) == > CPU_STATE_OFFLINE, > + 100000, 100); > > for (tries = 0; tries < 25; tries++) { Can we bumped up the tries to 100, so that we wait for 10ms before printing the warning message ? > cpu_status = smp_query_cpu_stopped(pcpu); > if (cpu_status == QCSS_STOPPED || > cpu_status == QCSS_HARDWARE_ERROR) > break; > - cpu_relax(); > + udelay(100); > } > } >
