Re: [PATCH kvm-unit-tests v1 4/6] s390x: smp: Create and use a non-waiting CPU stop

[Claudio Imbrenda <imbrenda@linux.ibm.com>]

On Thu,  3 Mar 2022 22:04:23 +0100
Eric Farman <farman@linux.ibm.com> wrote:

> When stopping a CPU, kvm-unit-tests serializes/waits for everything
> to finish, in order to get a consistent result whenever those
> functions are used.
> 
> But to test the SIGP STOP itself, these additional measures could
> mask other problems. For example, did the STOP work, or is the CPU
> still operating?
> 
> Let's create a non-waiting SIGP STOP and use it here, to ensure that
> the CPU is correctly stopped. A smp_cpu_stopped() call will still
> be used to see that the SIGP STOP has been processed, and the state
> of the CPU can be used to determine whether the test passes/fails.
> 
> Signed-off-by: Eric Farman <farman@linux.ibm.com>
> ---
>  lib/s390x/smp.c | 25 +++++++++++++++++++++++++
>  lib/s390x/smp.h |  1 +
>  s390x/smp.c     | 10 ++--------
>  3 files changed, 28 insertions(+), 8 deletions(-)
> 
> diff --git a/lib/s390x/smp.c b/lib/s390x/smp.c
> index 368d6add..84e536e8 100644
> --- a/lib/s390x/smp.c
> +++ b/lib/s390x/smp.c
> @@ -119,6 +119,31 @@ int smp_cpu_stop(uint16_t idx)
>  	return rc;
>  }
>  
> +/*
> + * Functionally equivalent to smp_cpu_stop(), but without the
> + * elements that wait/serialize matters itself.
> + * Used to see if KVM itself is serialized correctly.
> + */
> +int smp_cpu_stop_nowait(uint16_t idx)
> +{
> +	/* refuse to work on the boot CPU */
> +	if (idx == 0)
> +		return -1;
> +
> +	spin_lock(&lock);
> +
> +	/* Don't suppress a CC2 with sigp_retry() */
> +	if (smp_sigp(idx, SIGP_STOP, 0, NULL)) {
> +		spin_unlock(&lock);
> +		return -1;
> +	}
> +
> +	cpus[idx].active = false;
> +	spin_unlock(&lock);
> +
> +	return 0;
> +}
> +
>  int smp_cpu_stop_store_status(uint16_t idx)
>  {
>  	int rc;
> diff --git a/lib/s390x/smp.h b/lib/s390x/smp.h
> index 1e69a7de..bae03dfd 100644
> --- a/lib/s390x/smp.h
> +++ b/lib/s390x/smp.h
> @@ -44,6 +44,7 @@ bool smp_sense_running_status(uint16_t idx);
>  int smp_cpu_restart(uint16_t idx);
>  int smp_cpu_start(uint16_t idx, struct psw psw);
>  int smp_cpu_stop(uint16_t idx);
> +int smp_cpu_stop_nowait(uint16_t idx);
>  int smp_cpu_stop_store_status(uint16_t idx);
>  int smp_cpu_destroy(uint16_t idx);
>  int smp_cpu_setup(uint16_t idx, struct psw psw);
> diff --git a/s390x/smp.c b/s390x/smp.c
> index 50811bd0..11c2c673 100644
> --- a/s390x/smp.c
> +++ b/s390x/smp.c
> @@ -76,14 +76,8 @@ static void test_restart(void)
>  
>  static void test_stop(void)
>  {
> -	smp_cpu_stop(1);
> -	/*
> -	 * The smp library waits for the CPU to shut down, but let's
> -	 * also do it here, so we don't rely on the library
> -	 * implementation
> -	 */
> -	while (!smp_cpu_stopped(1)) {}
> -	report_pass("stop");
> +	smp_cpu_stop_nowait(1);

can it happen that the SIGP STOP order is accepted, but the target CPU
is still running (and not even busy)?

> +	report(smp_cpu_stopped(1), "stop");

e.g. can this ^ check race with the actual stopping of the CPU?

>  }
>  
>  static void test_stop_store_status(void)