Re: [PATCH 1/5] Fix soft-lookup in stop machine on secondary cpubring up

From: Igor Mammedov
Date: Fri May 11 2012 - 11:17:57 EST

On 05/11/2012 01:45 PM, Thomas Gleixner wrote:
On Wed, 9 May 2012, Igor Mammedov wrote:

When bringing up cpuX1, it could stall in start_secondary
before setting cpu_callin_mask for more than 5 sec. That forces
do_boot_cpu() to give up on waiting and go to error return path
printing messages:
pr_err("CPU%d: Stuck ??\n", cpuX1);
pr_err("CPU%d: Not responding.\n", cpuX1);
and native_cpu_up returns early with -EIO. However AP may continue
its boot process till it reaches check_tsc_sync_target(), where
it will wait for boot cpu to run cpu_up...=>check_tsc_sync_source.
That will never happen since cpu_up have returned with error before.

Now we need to note that cpuX1 is marked as active in smp_callin
before it stuck in check_tsc_sync_target. And when another cpuX2
is being onlined, start_secondary on it will call
-> smp_store_cpu_info
-> identify_secondary_cpu
-> mtrr_ap_init
-> set_mtrr_from_inactive_cpu
-> stop_machine_from_inactive_cpu
where it's going to schedule stop_machine work on all ACTIVE cpus
smdata.num_threads = num_active_cpus() + 1;
and wait till they all complete it before continuing. As was noted
before cpuX1 was marked as active but can't execute any work since
it's not completed initialization and stuck in check_tsc_sync_target.
As result system soft lockups in stop_machine_cpu_stop.

backtrace from reproducer:

PID: 3324 TASK: ffff88007c00ae20 CPU: other cpus COMMAND: "migration/1"
[exception RIP: stop_machine_cpu_stop+131]
#0 [ffff88007b4d7de8] cpu_stopper_thread at ffffffff810c66bd
#1 [ffff88007b4d7ee8] kthread at ffffffff8107871e
#2 [ffff88007b4d7f48] kernel_thread_helper at ffffffff8154af24

PID: 0 TASK: ffff88007c029710 CPU: 2 COMMAND: "swapper/2"
[exception RIP: check_tsc_sync_target+33]
#0 [ffff88007c025f30] start_secondary at ffffffff81539876

PID: 0 TASK: ffff88007c041710 CPU: 3 COMMAND: "swapper/3"
[exception RIP: stop_machine_cpu_stop+131]
#0 [ffff88007c04be50] stop_machine_from_inactive_cpu at ffffffff810c6b2f
#1 [ffff88007c04bee0] mtrr_ap_init at ffffffff8102e963
#2 [ffff88007c04bf10] identify_secondary_cpu at ffffffff81536799
#3 [ffff88007c04bf20] smp_store_cpu_info at ffffffff815396d5
#4 [ffff88007c04bf30] start_secondary at ffffffff81539800

Could be fixed by not marking being onlined cpu as active too early.

This explanation is completely useless. What's fixed by what. And is
it fixed or could it be fixed?
What's fixed:
above mentioned hang in stop_machine_from_inactive_cpu() because even if
a cpu failed to set cpu_callin_mask in time and boot cpu marked it as
not present + removed from some maps, with this move, a failed cpu won't
set cpu_active_mask before it completes check_tsc_sync_target().
And with patches [2,3,4]/5 it will not set cpu_active_mask at all
so making itself unavailable to the rest of kernel.

This also want's an explanation why moving the cpu_starting notifier
does not hurt any assumptions of the code which has notifiers
registered for CPU_STARTING.
I've checked in kernel users [sched, kvm, pmu] before moving it here.
It looked safe. However I might have missed something.

In fact your change can result in
CPU_ONLINE notifier being called _BEFORE_ CPU_STARTING. Do you really
think that's correct?
That's certainly is not correct, it asks for a barrier after
cpu_starting and before setting cpu_online_mask.

Aside of that your whole patch series tackles the wrong aspect.
patch series tries to prevent a failed to boot cpu wreck havoc on
running kernel. How wrong is that?
What should be fixed instead?

Why the heck do you need extra magic in check_tsc_sync_target() ?
Because it's plainly racy. patch 2/5 describes/fixes race condition in

If the booting CPU fails to set the callin map within 5 seconds then
it should not even reach check_tsc_sync_target() at all.
Why it shouldn't reach check_tsc_sync_target () at all. There is nothing
that prevents it and guaranties such behavior.
For example: it happens when kernel is running inside guest on overloaded host.
And it seems on baremetal as well:

And just for the record, the new CPU can run into the very same
timeout problem, when the boot CPU fails to set the callout mask.
Yes, it can.
I've tried to fix only what was reproducible on my test system, so I
haven't touched this.
That might result in panic in smp_callin():
panic("%s: CPU%d started up but did not get a callout!\n"

This whole stuff is a complete trainwreck already and I don't want to
see anything like your "fixing the symptoms" hackery near it, really.
Fixing slow to respond cpu might be not option, so we need to gracefully
abort failed cpu_online operation instead of hanging in stop_machine or
crashing in scheduler[].

This whole stuff needs a proper rewrite and not some more braindamaged
bandaids. And if we apply bandaids for the time being, then certainly
not bandaids like the mess you created.
Rewrite will need to deal with failed to boot in time cpu as well.
So if rewrite is not near completion, then maybe for a time being bandaids
would be needed.
Any ideas/suggestions for "right bandaids" instead of braindamaged ones?



To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at