kernel: cpu: Handle hotplug failure for state CPUHP_AP_IDLE_DEAD
PATCH 1/1 kernel: cpu: Handle hotplug failure for state CPUHP_AP_IDLE_DEAD
Once the tear down hotplug handler is run, cpu is dead and enters into CPUHP_AP_IDLE_DEAD state. Any callbacks that fail in the state machine with state < CPUHP_AP_IDLE must be treated as fatal as this could result into timer not beig migrated away from dead cpu and run into issues like work queue lock ups, sched_clock timer wrapping to zero as sched_clock_poll which is in the hrtimer base of cpu being hotplugged does not get migrated.
The function sched_clock_poll() updates the epoch_ns and epoch_cyc. If this function present in the hrtimer base of cpu being hotplugged doesn't migrate, there is no update on the epoch_ns and epoch_cyc. Subseqently, when sched_clock() is called, the non updated values of epoch_ns and epoch_cyc are obtained which looks like the timer wrapped around.
[ 8792.168842] pool 0: cpus=0 node=0 flags=0x0 nice=0 hung=6801s workers=2 manager: 4884 [ 8792.168862] pool 16: cpus=0-7 flags=0x4 nice=0 hung=0s workers=34 idle: 4482 1390 1394 1396 4492 5442 5447 5445 [ 0.017714] Modules linked in: wlan(O) [ 0.017733] CPU: 7 PID: 0 Comm: swapper/7 Tainted: G W O 4.9.37+ #1 [ 0.017746] task: ffffffc1b05c8080 task.stack: ffffffc1b05c4000
As seen, the time rolls over to 0 after 8792.
--- kernel/cpu.c | 1 + 1 file changed, 1 insertion(+)
kernel/cpu.c +1 -0
@@ -837,6 +837,7 @@ static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,for (; st->state > target; st->state--) { ret =cpuhp_invoke_callback (cpu, st->state, false,NULL ,NULL ); +BUG_ON (ret && st->state <CPUHP_AP_IDLE_DEAD );if (ret) { st->target = prev_state;undo_cpu_down (cpu, st);
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project
Quoted 7 lines · Prakruthi Deepak Heragu
Once the tear down hotplug handler is run, cpu is dead and enters into
CPUHP_AP_IDLE_DEADstate. Any callbacks that fail in the state machine with state <CPUHP_AP_IDLEmust be treated as fatal as this could result into timer not beig migrated away from dead cpu and run into issues like work queue lock ups,sched_clocktimer wrapping to zero assched_clock_pollwhich is in the hrtimer base of cpu being hotplugged does not get migrated.
BUG_ON() is the last resort when there is no other way out. And there is no reason to treat such a failure as fatal unconditionally.
Why would any of those callback fail at all? And if that ever happens, then we really can be smarter than just giving up.
Thanks,
tglx