KVM: x86: silence preempt warning on kvm_write_guest_time

KVM: x86: silence preempt warning on kvm_write_guest_time

[Marcelo Tosatti]

Glauber, Gerd,

See comments on
http://sourceforge.net/tracker2/?func=detail&aid=2627272&group_id=180599&atid=893831
regarding the "version" variable. Looks fishy.

From: Matt T. Yourst <yourst@users.sourceforge.net>

This issue just appeared in kvm-84 when running on 2.6.28.7 (x86-64)
with PREEMPT enabled.

We're getting syslog warnings like this many (but not all) times qemu
tells KVM to run the VCPU:

BUG: using smp_processor_id() in preemptible [00000000] code:
qemu-system-x86/28938
caller is kvm_arch_vcpu_ioctl_run+0x5d1/0xc70 [kvm]
Pid: 28938, comm: qemu-system-x86 2.6.28.7-mtyrel-64bit
Call Trace:
debug_smp_processor_id+0xf7/0x100
kvm_arch_vcpu_ioctl_run+0x5d1/0xc70 [kvm]
? __wake_up+0x4e/0x70
? wake_futex+0x27/0x40
kvm_vcpu_ioctl+0x2e9/0x5a0 [kvm]
enqueue_hrtimer+0x8a/0x110
_spin_unlock_irqrestore+0x27/0x50
vfs_ioctl+0x31/0xa0
do_vfs_ioctl+0x74/0x480
sys_futex+0xb4/0x140
sys_ioctl+0x99/0xa0
system_call_fastpath+0x16/0x1b

As it turns out, the call trace is messed up due to gcc's inlining, but
I isolated the problem anyway: kvm_write_guest_time() is being used in a
non-thread-safe manner on preemptable kernels.

Basically kvm_write_guest_time()'s body needs to be surrounded by
preempt_disable() and preempt_enable(), since the kernel won't let us
query any per-CPU data (indirectly using smp_processor_id()) without
preemption disabled. The attached patch fixes this issue by disabling
preemption inside kvm_write_guest_time().

[marcelo: surround only __get_cpu_var calls since the warning
is harmless]

Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 05d7be8..61aabb1 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -630,10 +630,12 @@ static void kvm_write_guest_time(struct kvm_vcpu *v)
 	if ((!vcpu->time_page))
 		return;
 
+	preempt_disable();
 	if (unlikely(vcpu->hv_clock_tsc_khz != __get_cpu_var(cpu_tsc_khz))) {
 		kvm_set_time_scale(__get_cpu_var(cpu_tsc_khz), &vcpu->hv_clock);
 		vcpu->hv_clock_tsc_khz = __get_cpu_var(cpu_tsc_khz);
 	}
+	preempt_enable();
 
 	/* Keep irq disabled to prevent changes to the clock */
 	local_irq_save(flags);

Re: KVM: x86: silence preempt warning on kvm_write_guest_time

[Gerd Hoffmann]

Marcelo Tosatti wrote:
> See comments on
> http://sourceforge.net/tracker2/?func=detail&aid=2627272&group_id=180599&atid=893831
> regarding the "version" variable. Looks fishy.

Fix looks correct (for both per-cpu access and the version stuff).

I think prempting the vcpu to another pcpu will force a kvmclock update,
so this should be handled just fine already.

cheers,
  Gerd

Re: KVM: x86: silence preempt warning on kvm_write_guest_time

[Avi Kivity]

Marcelo Tosatti wrote:
> Glauber, Gerd,
>
> See comments on
> http://sourceforge.net/tracker2/?func=detail&aid=2627272&group_id=180599&atid=893831
> regarding the "version" variable. Looks fishy.
>
> From: Matt T. Yourst <yourst@users.sourceforge.net>
>
> This issue just appeared in kvm-84 when running on 2.6.28.7 (x86-64)
> with PREEMPT enabled.
>
> We're getting syslog warnings like this many (but not all) times qemu
> tells KVM to run the VCPU:
>
> BUG: using smp_processor_id() in preemptible [00000000] code:
> qemu-system-x86/28938
> caller is kvm_arch_vcpu_ioctl_run+0x5d1/0xc70 [kvm]
> Pid: 28938, comm: qemu-system-x86 2.6.28.7-mtyrel-64bit
> Call Trace:
> debug_smp_processor_id+0xf7/0x100
> kvm_arch_vcpu_ioctl_run+0x5d1/0xc70 [kvm]
> ? __wake_up+0x4e/0x70
> ? wake_futex+0x27/0x40
> kvm_vcpu_ioctl+0x2e9/0x5a0 [kvm]
> enqueue_hrtimer+0x8a/0x110
> _spin_unlock_irqrestore+0x27/0x50
> vfs_ioctl+0x31/0xa0
> do_vfs_ioctl+0x74/0x480
> sys_futex+0xb4/0x140
> sys_ioctl+0x99/0xa0
> system_call_fastpath+0x16/0x1b
>
> As it turns out, the call trace is messed up due to gcc's inlining, but
> I isolated the problem anyway: kvm_write_guest_time() is being used in a
> non-thread-safe manner on preemptable kernels.
>
> Basically kvm_write_guest_time()'s body needs to be surrounded by
> preempt_disable() and preempt_enable(), since the kernel won't let us
> query any per-CPU data (indirectly using smp_processor_id()) without
> preemption disabled. The attached patch fixes this issue by disabling
> preemption inside kvm_write_guest_time().
>
> [marcelo: surround only __get_cpu_var calls since the warning
> is harmless]
>
>   

Applied, thanks.

-- 
error compiling committee.c: too many arguments to function