[PATCH 0/2] Improve DIAG 9c observability

[PATCH 0/2] Improve DIAG 9c observability

[Ciunas Bennett]

This series enhances observability for DIAG 9c directed yield operations
in s390 KVM by refactoring the handler and adding comprehensive tracing
support.

The first patch refactors __diag_time_slice_end_directed() to use a
single exit point with a result string variable. This consolidates the
control flow and eliminates code duplication in logging statements,
making the function more maintainable.

The second patch builds on this refactoring by adding a new tracepoint
kvm_s390_diag_9c that captures detailed information about directed yield
operations, including source/target VCPU IDs, physical CPU numbers, and
operation results. This enables better analysis of VCPU scheduling
behaviour and helps diagnose performance issues in virtualised s390
environments.

Together, these changes provide kernel developers and system
administrators with better tools to understand and debug VCPU scheduling
patterns related to directed yields.

Ciunas Bennett (2):
  s390/kvm: Refactor __diag_time_slice_end_directed for single exit
    point
  s390/kvm: Add tracepoint for DIAG 9c directed yield operations

 arch/s390/kvm/diag.c  | 19 +++++++++++--------
 arch/s390/kvm/trace.h | 30 ++++++++++++++++++++++++++++++
 2 files changed, 41 insertions(+), 8 deletions(-)

-- 
2.54.0

[PATCH 1/2] s390/kvm: Refactor __diag_time_slice_end_directed for single exit point

[Ciunas Bennett]

Refactor the DIAG 9c (directed yield) handler to use a unified exit path,
improving code maintainability and reducing duplication.
Changes:
 - Consolidate all exit paths to use a single 'out' label
 - Replace multiple VCPU_EVENT logging calls with one unified call
 - Introduce 'result' string variable to track operation outcome
 - Initialize tcpu_cpu to -1 for safe handling across all code paths
 - Ensure statistics updates occur before the common exit point

This refactoring maintains identical functionality while making the control
flow clearer and easier to maintain. All three possible outcomes (yield
forwarded, done, ignored) now converge at a single logging point

Signed-off-by: Ciunas Bennett <ciunas@linux.ibm.com>
---
 arch/s390/kvm/diag.c | 18 ++++++++++--------
 1 file changed, 10 insertions(+), 8 deletions(-)

diff --git a/arch/s390/kvm/diag.c b/arch/s390/kvm/diag.c
index d89d1c381522..85c84421b510 100644
--- a/arch/s390/kvm/diag.c
+++ b/arch/s390/kvm/diag.c
@@ -186,7 +186,8 @@ static int diag9c_forwarding_overrun(void)
 static int __diag_time_slice_end_directed(struct kvm_vcpu *vcpu)
 {
 	struct kvm_vcpu *tcpu;
-	int tcpu_cpu;
+	const char *result;
+	int tcpu_cpu = -1;
 	int tid;
 
 	tid = vcpu->run->s.regs.gprs[(vcpu->arch.sie_block->ipa & 0xf0) >> 4];
@@ -211,21 +212,22 @@ static int __diag_time_slice_end_directed(struct kvm_vcpu *vcpu)
 		if (!vcpu_is_preempted(tcpu_cpu))
 			goto no_yield;
 		smp_yield_cpu(tcpu_cpu);
-		VCPU_EVENT(vcpu, 5,
-			   "diag time slice end directed to %d: yield forwarded",
-			   tid);
 		vcpu->stat.diag_9c_forward++;
-		return 0;
+		result = "yield forwarded";
+		goto out;
 	}
 
 	if (kvm_vcpu_yield_to(tcpu) <= 0)
 		goto no_yield;
 
-	VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: done", tid);
-	return 0;
+	result = "done";
+	goto out;
 no_yield:
-	VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: ignored", tid);
 	vcpu->stat.diag_9c_ignored++;
+	result = "ignored";
+out:
+	VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: %s", tid,
+		   result);
 	return 0;
 }
 
-- 
2.54.0

[PATCH 2/2] s390/kvm: Add tracepoint for DIAG 9c directed yield operations

[Ciunas Bennett]

Add a new tracepoint kvm_s390_diag_9c to provide detailed observability
for directed yield operations. The tracepoint captures:
- Source and target VCPU IDs
- Current and target physical CPU numbers
- Operation result (done, ignored, yield forwarded)

This enables better analysis of VCPU scheduling behaviour and helps
diagnose performance issues related to directed yields in virtualised
s390 environments.

Signed-off-by: Ciunas Bennett <ciunas@linux.ibm.com>
---
 arch/s390/kvm/diag.c  |  1 +
 arch/s390/kvm/trace.h | 30 ++++++++++++++++++++++++++++++
 2 files changed, 31 insertions(+)

diff --git a/arch/s390/kvm/diag.c b/arch/s390/kvm/diag.c
index 85c84421b510..031ab6e5d6c4 100644
--- a/arch/s390/kvm/diag.c
+++ b/arch/s390/kvm/diag.c
@@ -228,6 +228,7 @@ static int __diag_time_slice_end_directed(struct kvm_vcpu *vcpu)
 out:
 	VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: %s", tid,
 		   result);
+	trace_kvm_s390_diag_9c(vcpu, tid, tcpu_cpu, result);
 	return 0;
 }
 
diff --git a/arch/s390/kvm/trace.h b/arch/s390/kvm/trace.h
index aa419eb6a0c8..2d6da21f590c 100644
--- a/arch/s390/kvm/trace.h
+++ b/arch/s390/kvm/trace.h
@@ -283,6 +283,36 @@ TRACE_EVENT(kvm_s390_handle_diag,
 			   __print_symbolic(__entry->code, diagnose_codes))
 	);
 
+TRACE_EVENT(kvm_s390_diag_9c,
+	    TP_PROTO(VCPU_PROTO_COMMON, int target_vcpu, int target_cpu,
+		     const char *result),
+	    TP_ARGS(VCPU_ARGS_COMMON, target_vcpu, target_cpu, result),
+
+	    TP_STRUCT__entry(
+		    VCPU_FIELD_COMMON
+		    __field(int, target_vcpu)
+		    __field(int, target_cpu)
+		    __field(int, current_cpu)
+		    __string(result, result)
+		    ),
+
+	    TP_fast_assign(
+		    VCPU_ASSIGN_COMMON
+		    __entry->target_vcpu = target_vcpu;
+		    __entry->target_cpu = target_cpu;
+		    __entry->current_cpu = smp_processor_id();
+		    __assign_str(result);
+		    ),
+
+	    VCPU_TP_PRINTK(
+		       "diag=9c vcpu=%d pcpu=%d target_vcpu=%d target_pcpu=%d result=%s",
+		       __entry->id,
+		       __entry->current_cpu,
+		       __entry->target_vcpu,
+		       __entry->target_cpu,
+		       __get_str(result))
+	);
+
 TRACE_EVENT(kvm_s390_handle_lctl,
 	    TP_PROTO(VCPU_PROTO_COMMON, int g, int reg1, int reg3, u64 addr),
 	    TP_ARGS(VCPU_ARGS_COMMON, g, reg1, reg3, addr),
-- 
2.54.0

Re: [PATCH 1/2] s390/kvm: Refactor __diag_time_slice_end_directed for single exit point

[Christian Borntraeger]

Am 19.06.26 um 12:13 schrieb Ciunas Bennett:
> Refactor the DIAG 9c (directed yield) handler to use a unified exit path,
> improving code maintainability and reducing duplication.
> Changes:
>   - Consolidate all exit paths to use a single 'out' label
>   - Replace multiple VCPU_EVENT logging calls with one unified call
>   - Introduce 'result' string variable to track operation outcome
>   - Initialize tcpu_cpu to -1 for safe handling across all code paths
>   - Ensure statistics updates occur before the common exit point
> 
> This refactoring maintains identical functionality while making the control
> flow clearer and easier to maintain. All three possible outcomes (yield
> forwarded, done, ignored) now converge at a single logging point
> 
> Signed-off-by: Ciunas Bennett <ciunas@linux.ibm.com>
> ---
>   arch/s390/kvm/diag.c | 18 ++++++++++--------
>   1 file changed, 10 insertions(+), 8 deletions(-)
> 
> diff --git a/arch/s390/kvm/diag.c b/arch/s390/kvm/diag.c
> index d89d1c381522..85c84421b510 100644
> --- a/arch/s390/kvm/diag.c
> +++ b/arch/s390/kvm/diag.c
> @@ -186,7 +186,8 @@ static int diag9c_forwarding_overrun(void)
>   static int __diag_time_slice_end_directed(struct kvm_vcpu *vcpu)
>   {
>   	struct kvm_vcpu *tcpu;
> -	int tcpu_cpu;
> +	const char *result;
> +	int tcpu_cpu = -1;
>   	int tid;
>   
>   	tid = vcpu->run->s.regs.gprs[(vcpu->arch.sie_block->ipa & 0xf0) >> 4];
> @@ -211,21 +212,22 @@ static int __diag_time_slice_end_directed(struct kvm_vcpu *vcpu)
>   		if (!vcpu_is_preempted(tcpu_cpu))
>   			goto no_yield;
>   		smp_yield_cpu(tcpu_cpu);
> -		VCPU_EVENT(vcpu, 5,
> -			   "diag time slice end directed to %d: yield forwarded",
> -			   tid);
>   		vcpu->stat.diag_9c_forward++;
> -		return 0;
> +		result = "yield forwarded";
> +		goto out;
>   	}
>   
>   	if (kvm_vcpu_yield_to(tcpu) <= 0)
>   		goto no_yield;
>   
> -	VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: done", tid);
> -	return 0;
> +	result = "done";
> +	goto out;
>   no_yield:
> -	VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: ignored", tid);
>   	vcpu->stat.diag_9c_ignored++;
> +	result = "ignored";
> +out:
> +	VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: %s", tid,
> +		   result);

result is a local variable that will go out of scope. You can not reference those
in the s390dbf feature as it will only store the string pointer and it will not
resolve the %s. So this wont work.

When sending a patch, please cc maintainers according to scripts/get_maintainers.pl

Re: [PATCH 1/2] s390/kvm: Refactor __diag_time_slice_end_directed for single exit point

[Ciunas Bennett]

On 19/06/2026 11:27, Christian Borntraeger wrote:
> 
> 
[ ... }
>> +        goto out;
>>       }
>>       if (kvm_vcpu_yield_to(tcpu) <= 0)
>>           goto no_yield;
>> -    VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: done", tid);
>> -    return 0;
>> +    result = "done";
>> +    goto out;
>>   no_yield:
>> -    VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: ignored", tid);
>>       vcpu->stat.diag_9c_ignored++;
>> +    result = "ignored";
>> +out:
>> +    VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: %s", tid,
>> +           result);
> 
> result is a local variable that will go out of scope. You can not reference those
> in the s390dbf feature as it will only store the string pointer and it will not
> resolve the %s. So this wont work.
> 
> When sending a patch, please cc maintainers according to scripts/get_maintainers.pl
> 

Hi Christian, your right I missed that.
Would something like this be acceptable?

/* set depending on state: e.g */
int result = 2;

VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: %s", tid,
        result == 0 ? "yield forwarded" :
        result == 1 ? "done" : "ignored");

Re: [PATCH 1/2] s390/kvm: Refactor __diag_time_slice_end_directed for single exit point

[Heiko Carstens]

On Fri, Jun 19, 2026 at 12:27:56PM +0200, Christian Borntraeger wrote:
> > +	const char *result;
...
> > +	result = "ignored";
> > +out:
> > +	VCPU_EVENT(vcpu, 5, "diag time slice end directed to %d: %s", tid,
> > +		   result);
> 
> result is a local variable that will go out of scope. You can not reference those
> in the s390dbf feature as it will only store the string pointer and it will not
> resolve the %s. So this wont work.

Why should this not work? The life time of result doesn't matter. The
_content_ of 'result', which is a pointer to ro.data section, where
the strings reside, is passed to the debug feature call. 'result' is
not required to resolve "%s" later.

But maybe I miss your point?