Re: [kvm-unit-tests PATCH v4 8/9] s390x: css: ssch/tsch with sense and interrupt

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

On 2019-12-12 13:26, Cornelia Huck wrote:
> On Wed, 11 Dec 2019 16:46:09 +0100
> Pierre Morel <pmorel@linux.ibm.com> wrote:
> 
>> When a channel is enabled we can start a SENSE command using the SSCH
> 
> s/SENSE/SENSE ID/
> 
> SENSE is for getting sense data after a unit check :)

Yes, thanks.

> 
>> instruction to recognize the control unit and device.
>>
>> This tests the success of SSCH, the I/O interruption and the TSCH
>> instructions.
>>
>> The test expects a device with a control unit type of 0xC0CA as the
>> first subchannel of the CSS.
>>
>> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com>
>> ---
>>   lib/s390x/css.h |  13 ++++
>>   s390x/css.c     | 175 ++++++++++++++++++++++++++++++++++++++++++++++++
>>   2 files changed, 188 insertions(+)
> 
>> +static void irq_io(void)
>> +{
>> +	int ret = 0;
>> +	char *flags;
>> +	int sid;
>> +
>> +	report_prefix_push("Interrupt");
>> +	if (lowcore->io_int_param != CSS_TEST_INT_PARAM) {
>> +		report(0, "Bad io_int_param: %x", lowcore->io_int_param);
>> +		report_prefix_pop();
>> +		return;
>> +	}
>> +	report_prefix_pop();
>> +
>> +	report_prefix_push("tsch");
>> +	sid = lowcore->subsys_id_word;
>> +	ret = tsch(sid, &irb);
>> +	switch (ret) {
>> +	case 1:
>> +		dump_irb(&irb);
>> +		flags = dump_scsw_flags(irb.scsw.ctrl);
>> +		report(0, "IRB scsw flags: %s", flags);
> 
> I guess that should only happen if the I/O interrupt was for another
> subchannel, and we only enable one subchannel, right?
> 
> Maybe log "I/O interrupt, but sch not status pending: <flags>"? (No
> idea how log the logged messages can be for kvm unit tests.)

Yes, the log message I had was not very useful at first sight.

> 
>> +		goto pop;
>> +	case 2:
>> +		report(0, "TSCH return unexpected CC 2");
> 
> s/return/returns/
> 
>> +		goto pop;
>> +	case 3:
>> +		report(0, "Subchannel %08x not operational", sid);
>> +		goto pop;
>> +	case 0:
>> +		/* Stay humble on success */
> 
> :)
> 
>> +		break;
>> +	}
>> +pop:
>> +	report_prefix_pop();
>> +}
>> +
>> +static int start_subchannel(int code, char *data, int count)
>> +{
>> +	int ret;
>> +	struct pmcw *p = &schib.pmcw;
>> +	struct orb *orb_p = &orb[0];
>> +
>> +	/* Verify that a test subchannel has been set */
>> +	if (!test_device_sid) {
>> +		report_skip("No device");
>> +		return 0;
>> +	}
>> +
>> +	if ((unsigned long)data >= 0x80000000UL) {
>> +		report(0, "Data above 2G! %p", data);
>> +		return 0;
>> +	}
>> +
>> +	/* Verify that the subchannel has been enabled */
>> +	ret = stsch(test_device_sid, &schib);
>> +	if (ret) {
>> +		report(0, "Err %d on stsch on sid %08x", ret, test_device_sid);
>> +		return 0;
>> +	}
>> +	if (!(p->flags & PMCW_ENABLE)) {
>> +		report_skip("Device (sid %08x) not enabled", test_device_sid);
>> +		return 0;
>> +	}
>> +
>> +	report_prefix_push("ssch");
>> +	/* Build the CCW chain with a single CCW */
>> +	ccw[0].code = code;
>> +	ccw[0].flags = 0; /* No flags need to be set */
>> +	ccw[0].count = count;
>> +	ccw[0].data_address = (int)(unsigned long)data;
>> +	orb_p->intparm = CSS_TEST_INT_PARAM;
>> +	orb_p->ctrl = ORB_F_INIT_IRQ|ORB_F_FORMAT|ORB_F_LPM_DFLT;
>> +	if ((unsigned long)&ccw[0] >= 0x80000000UL) {
>> +		report(0, "CCW above 2G! %016lx", (unsigned long)&ccw[0]);
> 
> Maybe check before filling out the ccw?

Yes. Also I wonder if we should not make sure the all kvm-test-text and 
data are under 2G by construct, because I am quite sure that this sort 
of tests will repeat all over the kvm-unit-test code.

Will provide a separate patch for this, in between just do as you said, 
it is the logical thing to do here.

> 
>> +		report_prefix_pop();
>> +		return 0;
>> +	}
>> +	orb_p->cpa = (unsigned int) (unsigned long)&ccw[0];
>> +
>> +	ret = ssch(test_device_sid, orb_p);
>> +	if (ret) {
>> +		report(0, "ssch cc=%d", ret);
>> +		report_prefix_pop();
>> +		return 0;
>> +	}
>> +	report_prefix_pop();
>> +	return 1;
>> +}
>> +
>> +/*
>> + * test_sense
>> + * Pre-requisits:
>> + * - We need the QEMU PONG device as the first recognized
>> + * - device by the enumeration.
>> + * - ./s390x-run s390x/css.elf -device ccw-pong,cu_type=0xc0ca
>> + */
>> +static void test_sense(void)
>> +{
>> +	int ret;
>> +
>> +	ret = register_io_int_func(irq_io);
>> +	if (ret) {
>> +		report(0, "Could not register IRQ handler");
>> +		goto unreg_cb;
>> +	}
>> +
>> +	enable_io_irq();
>> +	lowcore->io_int_param = 0;
>> +
>> +	ret = start_subchannel(CCW_CMD_SENSE_ID, buffer, sizeof(senseid));
>> +	if (!ret) {
>> +		report(0, "start_subchannel failed");
>> +		goto unreg_cb;
>> +	}
>> +
>> +	delay(100);
>> +	if (lowcore->io_int_param != CSS_TEST_INT_PARAM) {
>> +		report(0, "cu_type: expect 0x%08x, got 0x%08x",
>> +		       CSS_TEST_INT_PARAM, lowcore->io_int_param);
>> +		goto unreg_cb;
>> +	}
> 
> This still looks like that odd "delay and hope an interrupt has arrived
> in the mean time" pattern.

yes.

> 
> Also, doesn't the interrupt handler check for the intparm already?

Yes, if the interrupt fires.

> 
>> +
>> +	senseid.cu_type = buffer[2] | (buffer[1] << 8);
> 
> This still looks odd; why not have the ccw fill out the senseid
> structure directly?

Oh sorry, you already said and I forgot to modify this.
thanks

> 
>> +
>> +	/* Sense ID is non packed cut_type is at offset +1 byte */
> 
> I have trouble parsing this sentence...
> 
>> +	if (senseid.cu_type == PONG_CU)
>> +		report(1, "cu_type: expect 0x%04x got 0x%04x",
>> +		       PONG_CU_TYPE, senseid.cu_type);
>> +	else
>> +		report(0, "cu_type: expect 0x%04x got 0x%04x",
>> +		       PONG_CU_TYPE, senseid.cu_type);
> 
> Didn't you want to check for ff in the reserved field as well?

It was not intended as a check for SENSE_ID but for STSCH/READ.
But, while at this... why not.

Thanks for the review.
Regards,

Pierre


-- 
Pierre Morel
IBM Lab Boeblingen