On 11/26/2015 09:47 PM, Christian Borntraeger wrote:
> On 11/26/2015 05:17 PM, Tyler Baker wrote:
>> Hi Christian,
>>
>> The kernelci.org bot recently has been reporting kvm guest boot
>> failures[1] on various arm64 platforms in next-20151126. The bot
>> bisected[2] the failures to the commit in -next titled "KVM: Create
>> debugfs dir and stat files for each VM". I confirmed by reverting this
>> commit on top of next-20151126 it resolves the boot issue.
>>
>> In this test case the host and guest are booted with the same kernel.
>> The host is booted over nfs, installs qemu (qemu-system arm64 2.4.0),
>> and launches a guest. The host is booting fine, but when the guest is
>> launched it errors with "Failed to retrieve host CPU features!". I
>> checked the host logs, and found an "Unable to handle kernel paging
>> request" splat[3] which occurs when the guest is attempting to start.
>>
>> I scanned the patch in question but nothing obvious jumped out at me,
>> any thoughts?
>
> Not really.
> Do you have processing running that do read the files in /sys/kernel/debug/kvm/* ?
>
> If I read the arm oops message correctly it oopsed inside
> __srcu_read_lock. there is actually nothing in there that can oops,
> except the access to the preempt count. I am just guessing right now,
> but maybe the preempt variable is no longer available (as the process
> is gone). As long as a debugfs file is open, we hold a reference to
> the kvm, which holds a reference to the mm, so the mm might be killed
> after the process. But this is supposed to work, so maybe its something
> different. An objdump of __srcu_read_lock might help.
Hmm, the preempt thing is done in srcu_read_lock, but the crash is in
__srcu_read_lock. This function gets the srcu struct from mmu_notifier.c,
which must be present and is initialized during boot.
int __srcu_read_lock(struct srcu_struct *sp)
{
int idx;
idx = READ_ONCE(sp->completed) & 0x1;
__this_cpu_inc(sp->per_cpu_ref->c[idx]);
smp_mb(); /* B */ /* Avoid leaking the critical section. */
__this_cpu_inc(sp->per_cpu_ref->seq[idx]);
return idx;
}
Looking at the code I have no clue why the patch does make a difference.
Can you try to get an objdump -S for__Srcu_read_lock?
>
> I will drop it from my tree until we understand the problem
>
> Christian
>
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
