When booting a hyperthreaded system with the kernel parameter
'mitigations=auto,nosmt', the following warning occurs:
WARNING: CPU: 0 PID: 1 at drivers/xen/events/events_base.c:1112 unbind_from_irqhandler+0x4e/0x60
...
Hardware name: Xen HVM domU, BIOS 4.2.amazon 08/24/2006
...
Call Trace:
xen_uninit_lock_cpu+0x28/0x62
xen_hvm_cpu_die+0x21/0x30
takedown_cpu+0x9c/0xe0
? trace_suspend_resume+0x60/0x60
cpuhp_invoke_callback+0x9a/0x530
_cpu_up+0x11a/0x130
cpu_up+0x7e/0xc0
bringup_nonboot_cpus+0x48/0x50
smp_init+0x26/0x79
kernel_init_freeable+0xea/0x229
? rest_init+0xaa/0xaa
kernel_init+0xa/0x106
ret_from_fork+0x35/0x40
The secondary CPUs are not activated with the nosmt mitigations and only
the primary thread on each CPU core is used. In this situation,
xen_hvm_smp_prepare_cpus(), and more importantly xen_init_lock_cpu(), is
not called, so the lock_kicker_irq is not initialized for the secondary
CPUs. Let's fix this by exiting early in xen_uninit_lock_cpu() if the
irq is not set to avoid the warning from above for each secondary CPU.
Signed-off-by: Brian Masney <[email protected]>
---
arch/x86/xen/spinlock.c | 15 +++++++++++++++
1 file changed, 15 insertions(+)
diff --git a/arch/x86/xen/spinlock.c b/arch/x86/xen/spinlock.c
index 799f4eba0a62..4a052459a08e 100644
--- a/arch/x86/xen/spinlock.c
+++ b/arch/x86/xen/spinlock.c
@@ -93,9 +93,24 @@ void xen_init_lock_cpu(int cpu)
void xen_uninit_lock_cpu(int cpu)
{
+ int irq;
+
if (!xen_pvspin)
return;
+ /*
+ * When booting the kernel with 'mitigations=auto,nosmt', the secondary
+ * CPUs are not activated and only the primary thread on each CPU core
+ * is used. In this situation, xen_hvm_smp_prepare_cpus(), and more
+ * importantly xen_init_lock_cpu(), is not called, so the
+ * lock_kicker_irq is not initialized for the secondary CPUs. Let's
+ * exit early if the irq is not set to avoid a warning in the console
+ * log.
+ */
+ irq = per_cpu(lock_kicker_irq, cpu);
+ if (irq == -1)
+ return;
+
unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
per_cpu(lock_kicker_irq, cpu) = -1;
kfree(per_cpu(irq_name, cpu));
--
2.26.2
On Thu, Nov 05, 2020 at 07:35:29PM -0500, Brian Masney wrote:
> diff --git a/arch/x86/xen/spinlock.c b/arch/x86/xen/spinlock.c
> index 799f4eba0a62..4a052459a08e 100644
> --- a/arch/x86/xen/spinlock.c
> +++ b/arch/x86/xen/spinlock.c
> @@ -93,9 +93,24 @@ void xen_init_lock_cpu(int cpu)
>
> void xen_uninit_lock_cpu(int cpu)
> {
> + int irq;
> +
> if (!xen_pvspin)
> return;
>
> + /*
> + * When booting the kernel with 'mitigations=auto,nosmt', the secondary
> + * CPUs are not activated and only the primary thread on each CPU core
> + * is used. In this situation, xen_hvm_smp_prepare_cpus(), and more
> + * importantly xen_init_lock_cpu(), is not called, so the
> + * lock_kicker_irq is not initialized for the secondary CPUs. Let's
> + * exit early if the irq is not set to avoid a warning in the console
> + * log.
> + */
> + irq = per_cpu(lock_kicker_irq, cpu);
> + if (irq == -1)
> + return;
> +
> unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
As soon as I saw this on lore, I saw that I should have passed the irq
variable to unbind_from_irqhandler() rather than doing another per_cpu()
lookup. I'll wait for feedback about the general approach before posting
a v2.
Brian
On 11/5/20 7:47 PM, Brian Masney wrote:
> On Thu, Nov 05, 2020 at 07:35:29PM -0500, Brian Masney wrote:
>> diff --git a/arch/x86/xen/spinlock.c b/arch/x86/xen/spinlock.c
>> index 799f4eba0a62..4a052459a08e 100644
>> --- a/arch/x86/xen/spinlock.c
>> +++ b/arch/x86/xen/spinlock.c
>> @@ -93,9 +93,24 @@ void xen_init_lock_cpu(int cpu)
>>
>> void xen_uninit_lock_cpu(int cpu)
>> {
>> + int irq;
>> +
>> if (!xen_pvspin)
>> return;
>>
>> + /*
>> + * When booting the kernel with 'mitigations=auto,nosmt', the secondary
>> + * CPUs are not activated and only the primary thread on each CPU core
>> + * is used. In this situation, xen_hvm_smp_prepare_cpus(), and more
>> + * importantly xen_init_lock_cpu(), is not called, so the
>> + * lock_kicker_irq is not initialized for the secondary CPUs. Let's
>> + * exit early if the irq is not set to avoid a warning in the console
>> + * log.
>> + */
>> + irq = per_cpu(lock_kicker_irq, cpu);
>> + if (irq == -1)
>> + return;
>> +
>> unbind_from_irqhandler(per_cpu(lock_kicker_irq, cpu), NULL);
> As soon as I saw this on lore, I saw that I should have passed the irq
> variable to unbind_from_irqhandler() rather than doing another per_cpu()
> lookup. I'll wait for feedback about the general approach before posting
> a v2.
This looks good. I'd shorten the comment though: your commit message already describes the scenario. And change the subject to something like "Don't unbind uninitialized lock_kicker_irq".
-boris