Re: [kvm-unit-tests PATCH v4 7/9] s390x: css: msch, enable test

[Pierre Morel <pmorel@linux.ibm.com>]

On 2019-12-12 15:10, Cornelia Huck wrote:
> On Thu, 12 Dec 2019 15:01:07 +0100
> Pierre Morel <pmorel@linux.ibm.com> wrote:
> 
>> On 2019-12-12 13:01, Cornelia Huck wrote:
>>> On Wed, 11 Dec 2019 16:46:08 +0100
>>> Pierre Morel <pmorel@linux.ibm.com> wrote:
>>>    
>>>> A second step when testing the channel subsystem is to prepare a channel
>>>> for use.
>>>> This includes:
>>>> - Get the current SubCHannel Information Block (SCHIB) using STSCH
>>>> - Update it in memory to set the ENABLE bit
>>>> - Tell the CSS that the SCHIB has been modified using MSCH
>>>> - Get the SCHIB from the CSS again to verify that the subchannel is
>>>>     enabled.
>>>>
>>>> This tests the success of the MSCH instruction by enabling a channel.
>>>>
>>>> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com>
>>>> ---
>>>>    s390x/css.c | 65 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>>>>    1 file changed, 65 insertions(+)
> 
>>>> +	/* Read the SCHIB for this subchannel */
>>>> +	cc = stsch(test_device_sid, &schib);
>>>> +	if (cc) {
>>>> +		report(0, "stsch cc=%d", cc);
>>>> +		return;
>>>> +	}
>>>> +
>>>> +	/* Update the SCHIB to enable the channel */
>>>> +	pmcw->flags |= PMCW_ENABLE;
>>>> +
>>>> +	/* Tell the CSS we want to modify the subchannel */
>>>> +	cc = msch(test_device_sid, &schib);
>>>> +	if (cc) {
>>>> +		/*
>>>> +		 * If the subchannel is status pending or
>>>> +		 * if a function is in progress,
>>>> +		 * we consider both cases as errors.
>>>> +		 */
>>>> +		report(0, "msch cc=%d", cc);
>>>> +		return;
>>>> +	}
>>>> +
>>>> +	/*
>>>> +	 * Read the SCHIB again to verify the enablement
>>>> +	 * insert a little delay and try 5 times.
>>>> +	 */
>>>> +	do {
>>>> +		cc = stsch(test_device_sid, &schib);
>>>> +		if (cc) {
>>>> +			report(0, "stsch cc=%d", cc);
>>>> +			return;
>>>> +		}
>>>> +		delay(10);
>>>
>>> That's just a short delay to avoid a busy loop, right? msch should be
>>> immediate,
>>
>> Thought you told to me that it may not be immediate in zVM did I
>> misunderstand?
> 
> Maybe I have been confusing... what I'm referring to is this
> programming note for msch:
> 
> "It is recommended that the program inspect the
> contents of the subchannel by subsequently
> issuing STORE SUBCHANNEL when MODIFY
> SUBCHANNEL sets condition code 0. Use of
> STORE SUBCHANNEL is a method for deter-
> mining if the designated subchannel was
> changed or not. Failure to inspect the subchan-
> nel following the setting of condition code 0 by
> MODIFY SUBCHANNEL may result in conditions
> that the program does not expect to occur."
> 
> That's exactly what we had to do under z/VM back then: do the msch,
> check via stsch, redo the msch if needed, check again via stsch. It
> usually worked with the second msch the latest.

OK, I understand, then it is a bug in zVM that this test could enlighten.

I think we should keep it so, it allows to recognize 3 cases (after I 
change to test ENABLE in the loop as I said I will):
- immediate ENABLE
- asynchrone ENABLE
- failure to ENABLE


> 
>>
>>> and you probably should not delay on success?
>>
>> yes, it is not optimized, I can test PMCW_ENABLE in the loop this way we
>> can see if, in the zVM case we need to do retries or not.
>>
>>
>>>    
>>>> +	} while (!(pmcw->flags & PMCW_ENABLE) && count++ < 5);
>>>
>>> How is this supposed to work? Doesn't the stsch overwrite the control
>>> block again, so you need to re-set the enable bit before you retry?
>>
>> I do not think so, there is no msch() in the loop.
>> Do I miss something?
> 
> Well, _I_ missed that the msch() was missing :) You need it (see above);
> just waiting and re-doing the stsch is useless, as msch is a
> synchronous instruction which has finished its processing after the cc
> has been set.
> 

Since kvm-unit-test is a test system, not an OS so I think that here we 
have one more point to leverage the enable function:
- We need to test the enable (what I did (partially))
- We need the enable to work (your proposition) to further test the I/O

OK, I rework this part with your comment in mind.

Thanks
Pierre


-- 
Pierre Morel
IBM Lab Boeblingen