Re: [PATCH v2 08/10] s390x: smp: Wait for sigp completion

[David Hildenbrand <david@redhat.com>]

On 29.04.20 11:37, Janosch Frank wrote:
> On 4/29/20 11:06 AM, David Hildenbrand wrote:
>> On 29.04.20 10:57, Janosch Frank wrote:
>>> On 4/24/20 1:40 PM, Janosch Frank wrote:
>>>> On 4/24/20 12:11 PM, David Hildenbrand wrote:
>>>>> On 23.04.20 11:10, Janosch Frank wrote:
>>>>>> Sigp orders are not necessarily finished when the processor finished
>>>>>> the sigp instruction. We need to poll if the order has been finished
>>>>>> before we continue.
>>>>>>
>>>>>> For (re)start and stop we already use sigp sense running and sigp
>>>>>> sense loops. But we still lack completion checks for stop and store
>>>>>> status, as well as the cpu resets.
>>>>>>
>>>>>> Let's add them.
>>>>>>
>>>>>> Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
>>>>>> Reviewed-by: Cornelia Huck <cohuck@redhat.com>
>>>>>> ---
>>>>>>  lib/s390x/smp.c | 8 ++++++++
>>>>>>  lib/s390x/smp.h | 1 +
>>>>>>  s390x/smp.c     | 4 ++++
>>>>>>  3 files changed, 13 insertions(+)
>>>>>>
>>>>>> diff --git a/lib/s390x/smp.c b/lib/s390x/smp.c
>>>>>> index 6ef0335..2555bf4 100644
>>>>>> --- a/lib/s390x/smp.c
>>>>>> +++ b/lib/s390x/smp.c
>>>>>> @@ -154,6 +154,14 @@ int smp_cpu_start(uint16_t addr, struct psw psw)
>>>>>>  	return rc;
>>>>>>  }
>>>>>>  
>>>>>> +void smp_cpu_wait_for_completion(uint16_t addr)
>>>>>> +{
>>>>>> +	uint32_t status;
>>>>>> +
>>>>>> +	/* Loops when cc == 2, i.e. when the cpu is busy with a sigp order */
>>>>>> +	sigp_retry(1, SIGP_SENSE, 0, &status);
>>>>>> +}
>>>>>> +
>>>>>>  int smp_cpu_destroy(uint16_t addr)
>>>>>>  {
>>>>>>  	struct cpu *cpu;
>>>>>> diff --git a/lib/s390x/smp.h b/lib/s390x/smp.h
>>>>>> index ce63a89..a8b98c0 100644
>>>>>> --- a/lib/s390x/smp.h
>>>>>> +++ b/lib/s390x/smp.h
>>>>>> @@ -45,6 +45,7 @@ int smp_cpu_restart(uint16_t addr);
>>>>>>  int smp_cpu_start(uint16_t addr, struct psw psw);
>>>>>>  int smp_cpu_stop(uint16_t addr);
>>>>>>  int smp_cpu_stop_store_status(uint16_t addr);
>>>>>> +void smp_cpu_wait_for_completion(uint16_t addr);
>>>>>>  int smp_cpu_destroy(uint16_t addr);
>>>>>>  int smp_cpu_setup(uint16_t addr, struct psw psw);
>>>>>>  void smp_teardown(void);
>>>>>> diff --git a/s390x/smp.c b/s390x/smp.c
>>>>>> index 7462211..48321f4 100644
>>>>>> --- a/s390x/smp.c
>>>>>> +++ b/s390x/smp.c
>>>>>> @@ -75,6 +75,7 @@ static void test_stop_store_status(void)
>>>>>>  	lc->prefix_sa = 0;
>>>>>>  	lc->grs_sa[15] = 0;
>>>>>>  	smp_cpu_stop_store_status(1);
>>>>>> +	smp_cpu_wait_for_completion(1);
>>>>>>  	mb();
>>>>>>  	report(lc->prefix_sa == (uint32_t)(uintptr_t)cpu->lowcore, "prefix");
>>>>>>  	report(lc->grs_sa[15], "stack");
>>>>>> @@ -85,6 +86,7 @@ static void test_stop_store_status(void)
>>>>>>  	lc->prefix_sa = 0;
>>>>>>  	lc->grs_sa[15] = 0;
>>>>>>  	smp_cpu_stop_store_status(1);
>>>>>
>>>>> Just curious: Would it make sense to add that inside
>>>>> smp_cpu_stop_store_status() instead?
>>>>>
>>>>
>>>> I think so, we also wait for stop and start to finish, so why not for
>>>> this order code.
>>>>
>>>
>>> I've moved the waiting into the smp library and now the prefix check for
>>> stop and store status fails every so often if executed repeatedly.
>>>
>>> I've tried making the lc ptr volatile, a print of the prefix before the
>>> report seems to fix the issue, a print after the report still shows the
>>> issue but according to the print both values are the same.
>>>
>>> I'm currently at a loss...
>>
>> Are you missing a barrier() somewhere?
>>
> 
> Maybe, but the question is where?
> 
> There's already one before the report:
> smp_cpu_stop_store_status(1);
> mb();
> report(lc->prefix_sa == (uint32_t)(uintptr_t)cpu->lowcore, "prefix");

The issue here is:

SIGP_SENSE is always handled in the kernel for KVM. Meaning, it will
complete even before the target CPU executed the stop and store (in QEMU).

Reading the PoP:

"One of the following conditions exists at the
addressed CPU: ... A previously issued stop-
and-store-status ... has been accepted by the
addressed CPU, and execution of the func-
tion requested by the order has not yet been
completed.

"If the currently specified order is sense ... then the order
is rejected, and condition code 2 is set."

So, in case of KVM, SENSE does not wait for completion of the previous
order. I remember that was a performance improvements, because we wanted
to avoid going to user space just to sense if another CPU is running.
(and I remember that the documentation was inconsistent)

Let me guess, under TCG it works all the time?

-- 
Thanks,

David / dhildenb