Re: [PATCH] Revert "qemu-kvm: Bring qemu_init_vcpu back home"

[Jan Kiszka <jan.kiszka@siemens.com>]

On 2011-02-23 15:38, Ryousei Takano wrote:
> This reverts commit c65827770b5e7282b5fb9da29c1de85ed08e816e

Reasoning would belong here, not below the separator.

> 
> Signed-off-by: Ryousei Takano <takano-ryousei@aist.go.jp>
> ---
> I am trying to run OpenMP (i.e., multi-threaded) programs using Intel
> Math Kernel Library (MKL) on a guest OS.  In the qemu-kvm-0.12.5,
> my program works fine.  In the qemu-kvm-0.13.0, however, it does not.

What about 0.14?

> The Linux kernel is 2.6.32.28.  The outputs of 'ps -L' are shown as follows:
> 
> qemu-kvm-0.12.5:
>  PID   LWP TTY          TIME CMD
>  2731  2731 pts/0    00:00:00 bash
>  2799  2799 pts/0    00:00:19 a.out
>  2799  2803 pts/0    00:00:00 a.out
>  2799  2804 pts/0    00:00:04 a.out   <-
>  2799  2805 pts/0    00:00:03 a.out   <- these threads are working properly.
>  2799  2806 pts/0    00:00:04 a.out   <-
>  2807  2807 pts/0    00:00:00 ps
> 
> qemu-0.13.0:
>  PID   LWP TTY          TIME CMD
>  2725  2725 pts/0    00:00:00 bash
>  2748  2748 pts/0    00:00:19 a.out
>  2748  2752 pts/0    00:00:00 a.out
>  2748  2753 pts/0    00:00:00 a.out   <-
>  2748  2754 pts/0    00:00:00 a.out   <- these threads are blocked by
> pthread_cond_wait(3).
>  2748  2755 pts/0    00:00:00 a.out   <-
>  2756  2756 pts/0    00:00:00 ps
> 
> This problem seems to be introduced by the commit
> c65827770b5e7282b5fb9da29c1de85ed08e816e.
> The attached revert patch solves it.  But I do not understand what is
> happened;-)
> Any comments and suggestions will be appreciated.
> 
> Regards,
> Ryousei
> 
>  hw/pc.c              |    6 ++++++
>  target-i386/helper.c |    2 --
>  2 files changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/hw/pc.c b/hw/pc.c
> index ec97181..577abe7 100644
> --- a/hw/pc.c
> +++ b/hw/pc.c
> @@ -937,6 +937,12 @@ CPUState *pc_new_cpu(const char *cpu_model)
>      }
>      qemu_register_reset(pc_cpu_reset, env);
>      pc_cpu_reset(env);
> +
> +    /* kvm needs this to run after the apic is initialized. Otherwise,
> +     * it can access invalid state and crash.
> +     */
> +    qemu_init_vcpu(env);
> +
>      return env;
>  }
> 
> diff --git a/target-i386/helper.c b/target-i386/helper.c
> index 86202e6..69cda5e 100644
> --- a/target-i386/helper.c
> +++ b/target-i386/helper.c
> @@ -1224,8 +1224,6 @@ CPUX86State *cpu_x86_init(const char *cpu_model)
>      }
>      mce_init(env);
> 
> -    qemu_init_vcpu(env);
> -
>      return env;
>  }
> 

I don't believe yet that the patch is the reason, specifically as it
targets the setup phase (except for CPU hotplug, but that's broken for
ages), not the operational phase of the VM.

What's your qemu-kvm command line?

Also, please attach gdb to the stuck VM and print the backtraces of all
threads ("thread apply all bt").

Thanks,
Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux