While playing with kvm cpu hot-plug, I've probably stumbled on general
kernel bug. So I'm lookng for advice on approach to fix it.
When kvm guest uses kvmclock, it may hang on cpu hot-plug at
BSP:
smp_apic_timer_interrupt
...
-> do_timer
-> update_wall_time
and a being on-lined CPU at waiting on sync with BSP at:
start_secondary:
while (!cpumask_test_cpu(smp_processor_id(), cpu_active_mask))
cpu_relax();
because of unusually big value returned by clock->read(clock) in
update_wall_time.
This happens due to overflow in pvclock_get_nsec_offset
u64 delta = tsc - shadow->tsc_timestamp;
when shadow->tsc_timestamp is bigger than tsc.
And since pvclock_clocksource_read remembers and returns largest
value of any clock that ever was returned, clock for affected guest
virtually freezes.
Overflow happens due to reading undefined values from uninitialized
per_cpu variable hv_clock. In case of cpu hot-plug, clock is read at
least once on being on-lined cpu before it is initialized for this
cpu:
start_secondary
|-> smp_callin
| -> smp_store_cpu_info
| -> identify_secondary_cpu
| -> mtrr_ap_init
| -> mtrr_restore
| -> stop_machine_from_inactive_cpu
| -> queue_stop_cpus_work
| ...
| -> kvm_clock_read
|...
|x86_cpuinit.setup_percpu_clockev
full call chain is below:
[ 0.002999] [<ffffffff8102fc74>] pvclock_clocksource_read+0x9f/0x275
[ 0.002999] [<ffffffff81071f9e>] ? check_preempt_wakeup+0x11c/0x1c2
[ 0.002999] [<ffffffff8102f05d>] kvm_clock_read+0x21/0xd3
[ 0.002999] [<ffffffff810154b1>] sched_clock+0x9/0xd
[ 0.002999] [<ffffffff81070509>] sched_clock_local+0x12/0x75
[ 0.002999] [<ffffffff8107065e>] sched_clock_cpu+0x84/0xc6
[ 0.002999] [<ffffffff8106bfd3>] update_rq_clock+0x28/0x108
[ 0.002999] [<ffffffff8106c15e>] enqueue_task+0x1d/0x64
[ 0.002999] [<ffffffff8106c500>] activate_task+0x22/0x24
[ 0.002999] [<ffffffff8106c90c>] ttwu_do_activate.constprop.39+0x32/0x61
[ 0.002999] [<ffffffff8106cd0c>] try_to_wake_up+0x17e/0x1e1
[ 0.002999] [<ffffffff8106cd98>] wake_up_process+0x15/0x17
[ 0.002999] [<ffffffff8109ebbf>] cpu_stop_queue_work+0x3d/0x5f
[ 0.002999] [<ffffffff8109ec6c>] queue_stop_cpus_work+0x8b/0xb6
[ 0.002999] [<ffffffff8109f0e3>] stop_machine_from_inactive_cpu+0xb4/0xed
[ 0.002999] [<ffffffff81023896>] ? mtrr_restore+0x4a/0x4a
[ 0.002999] [<ffffffff81023eb3>] mtrr_ap_init+0x5a/0x5c
[ 0.002999] [<ffffffff814b95f7>] identify_secondary_cpu+0x19/0x1b
[ 0.002999] [<ffffffff814bc0d3>] smp_store_cpu_info+0x3c/0x3e
[ 0.002999] [<ffffffff814bc4ed>] start_secondary+0x122/0x263
Looking at native_smp_prepare_cpus in arch/x86/kernel/smpboot.c
I see that it unconditionally calls set_mtrr_aps_delayed_init which
in turn effectively converts calls to mtrr_ap_init to nop ops.
And later call to native_smp_cpus_done -> mtrr_aps_init completes
mtrr initialization.
The same pattern might be noticed in suspend/hibernate handlers, that
call enable_nonboot_cpus.
enable_nonboot_cpus
-> arch_enable_nonboot_cpus_begin
-> set_mtrr_aps_delayed_init
-> boot secondary cpus
-> arch_enable_nonboot_cpus_end
-> mtrr_aps_init
So question is if the calling mtrr_ap_init so early in
identify_secondary_cpu is really necessary?
From current code it looks like it is never called at normal smp
boot/resume time. And only path that triggers it is cpu hot-plug one.
I see following possible solutions:
1. Could the call to mtrr_ap_init be just moved from identify_secondary_cpu
to start_secondary right after x86_cpuinit.setup_percpu_clockev call?
This will prevent an access to uninitialized per_cpu clock. Or move
x86_cpuinit.setup_percpu_clockev before smp_callin?
2. Another way to prevent access to uninitialized per_cpu clock is to
introduce hook.early_percpu_clock_init that will be called before
smp_callin or right before mtrr_ap_init.
I hope to see opinions about this matter from a more experienced people than me.
--
Thanks,
Igor
On Thu, Feb 02, 2012 at 12:43:00PM +0100, Igor Mammedov wrote:
> While playing with kvm cpu hot-plug, I've probably stumbled on general
> kernel bug. So I'm lookng for advice on approach to fix it.
>
> When kvm guest uses kvmclock, it may hang on cpu hot-plug at
>
> BSP:
> smp_apic_timer_interrupt
> ...
> -> do_timer
> -> update_wall_time
>
> and a being on-lined CPU at waiting on sync with BSP at:
> start_secondary:
> while (!cpumask_test_cpu(smp_processor_id(), cpu_active_mask))
> cpu_relax();
>
> because of unusually big value returned by clock->read(clock) in
> update_wall_time.
> This happens due to overflow in pvclock_get_nsec_offset
>
> u64 delta = tsc - shadow->tsc_timestamp;
>
> when shadow->tsc_timestamp is bigger than tsc.
> And since pvclock_clocksource_read remembers and returns largest
> value of any clock that ever was returned, clock for affected guest
> virtually freezes.
>
> Overflow happens due to reading undefined values from uninitialized
> per_cpu variable hv_clock. In case of cpu hot-plug, clock is read at
> least once on being on-lined cpu before it is initialized for this
> cpu:
>
> start_secondary
> |-> smp_callin
> | -> smp_store_cpu_info
> | -> identify_secondary_cpu
> | -> mtrr_ap_init
> | -> mtrr_restore
> | -> stop_machine_from_inactive_cpu
> | -> queue_stop_cpus_work
> | ...
> | -> kvm_clock_read
> |...
> |x86_cpuinit.setup_percpu_clockev
>
> full call chain is below:
> [ 0.002999] [<ffffffff8102fc74>] pvclock_clocksource_read+0x9f/0x275
> [ 0.002999] [<ffffffff81071f9e>] ? check_preempt_wakeup+0x11c/0x1c2
> [ 0.002999] [<ffffffff8102f05d>] kvm_clock_read+0x21/0xd3
> [ 0.002999] [<ffffffff810154b1>] sched_clock+0x9/0xd
> [ 0.002999] [<ffffffff81070509>] sched_clock_local+0x12/0x75
> [ 0.002999] [<ffffffff8107065e>] sched_clock_cpu+0x84/0xc6
> [ 0.002999] [<ffffffff8106bfd3>] update_rq_clock+0x28/0x108
> [ 0.002999] [<ffffffff8106c15e>] enqueue_task+0x1d/0x64
> [ 0.002999] [<ffffffff8106c500>] activate_task+0x22/0x24
> [ 0.002999] [<ffffffff8106c90c>] ttwu_do_activate.constprop.39+0x32/0x61
> [ 0.002999] [<ffffffff8106cd0c>] try_to_wake_up+0x17e/0x1e1
> [ 0.002999] [<ffffffff8106cd98>] wake_up_process+0x15/0x17
> [ 0.002999] [<ffffffff8109ebbf>] cpu_stop_queue_work+0x3d/0x5f
> [ 0.002999] [<ffffffff8109ec6c>] queue_stop_cpus_work+0x8b/0xb6
> [ 0.002999] [<ffffffff8109f0e3>] stop_machine_from_inactive_cpu+0xb4/0xed
> [ 0.002999] [<ffffffff81023896>] ? mtrr_restore+0x4a/0x4a
> [ 0.002999] [<ffffffff81023eb3>] mtrr_ap_init+0x5a/0x5c
> [ 0.002999] [<ffffffff814b95f7>] identify_secondary_cpu+0x19/0x1b
> [ 0.002999] [<ffffffff814bc0d3>] smp_store_cpu_info+0x3c/0x3e
> [ 0.002999] [<ffffffff814bc4ed>] start_secondary+0x122/0x263
>
> Looking at native_smp_prepare_cpus in arch/x86/kernel/smpboot.c
> I see that it unconditionally calls set_mtrr_aps_delayed_init which
> in turn effectively converts calls to mtrr_ap_init to nop ops.
> And later call to native_smp_cpus_done -> mtrr_aps_init completes
> mtrr initialization.
>
> The same pattern might be noticed in suspend/hibernate handlers, that
> call enable_nonboot_cpus.
>
> enable_nonboot_cpus
> -> arch_enable_nonboot_cpus_begin
> -> set_mtrr_aps_delayed_init
>
> -> boot secondary cpus
>
> -> arch_enable_nonboot_cpus_end
> -> mtrr_aps_init
>
> So question is if the calling mtrr_ap_init so early in
> identify_secondary_cpu is really necessary?
> From current code it looks like it is never called at normal smp
> boot/resume time. And only path that triggers it is cpu hot-plug one.
>
> I see following possible solutions:
>
> 1. Could the call to mtrr_ap_init be just moved from identify_secondary_cpu
> to start_secondary right after x86_cpuinit.setup_percpu_clockev call?
>
> This will prevent an access to uninitialized per_cpu clock. Or move
> x86_cpuinit.setup_percpu_clockev before smp_callin?
>
> 2. Another way to prevent access to uninitialized per_cpu clock is to
> introduce hook.early_percpu_clock_init that will be called before
> smp_callin or right before mtrr_ap_init.
>
> I hope to see opinions about this matter from a more experienced people than me.
It would be more robust to introduce x86_cpuinit.early_percpu_clock_init
and use it for KVM, replacing setup_percpu_clockev hook.
setup_percpu_clockev initializes the APIC timer, but clearly sched_clock
should be functional before that, for example for ftrace (in addition to
sched_clock from scheduler as in the trace above).