Re: [PATCH net-next 3/4] dpll: netlink/core: add support for pin-dpll signal phase offset/adjust

[Jiri Pirko <jiri@resnulli.us>]

Tue, Oct 03, 2023 at 04:29:43PM CEST, arkadiusz.kubalewski@intel.com wrote:
>>From: Jiri Pirko <jiri@resnulli.us>
>>Sent: Tuesday, October 3, 2023 8:32 AM
>>
>>Tue, Oct 03, 2023 at 01:03:00AM CEST, arkadiusz.kubalewski@intel.com wrote:
>>>>From: Jiri Pirko <jiri@resnulli.us>
>>>>Sent: Monday, October 2, 2023 5:04 PM
>>>>
>>>>Mon, Oct 02, 2023 at 04:32:30PM CEST, arkadiusz.kubalewski@intel.com
>>>>wrote:
>>>>>>From: Vadim Fedorenko <vadim.fedorenko@linux.dev>
>>>>>>Sent: Wednesday, September 27, 2023 8:09 PM
>>>>>>
>>>>>>On 27/09/2023 10:24, Arkadiusz Kubalewski wrote:
>>>>>>> Add callback op (get) for pin-dpll phase-offset measurment.
>>>>>>> Add callback ops (get/set) for pin signal phase adjustment.
>>>>>>> Add min and max phase adjustment values to pin proprties.
>>>>>>> Invoke get callbacks when filling up the pin details to provide user
>>>>>>> with phase related attribute values.
>>>>>>> Invoke phase-adjust set callback when phase-adjust value is provided
>>>>>>> for
>>>>>>> pin-set request.
>>>>>>>
>>>>>>> Signed-off-by: Arkadiusz Kubalewski <arkadiusz.kubalewski@intel.com>
>>>>>>
>>>>>>[...]
>>>>>>
>>>>>>> +static int
>>>>>>> +dpll_pin_phase_adj_set(struct dpll_pin *pin, struct nlattr
>>>>>>> *phase_adj_attr,
>>>>>>> +		       struct netlink_ext_ack *extack)
>>>>>>> +{
>>>>>>> +	struct dpll_pin_ref *ref;
>>>>>>> +	unsigned long i;
>>>>>>> +	s32 phase_adj;
>>>>>>> +	int ret;
>>>>>>> +
>>>>>>> +	phase_adj = nla_get_s32(phase_adj_attr);
>>>>>>> +	if (phase_adj > pin->prop->phase_range.max ||
>>>>>>> +	    phase_adj < pin->prop->phase_range.min) {
>>>>>>> +		NL_SET_ERR_MSG(extack, "phase adjust value not
>>>>>>> supported");
>>>>>>> +		return -EINVAL;
>>>>>>> +	}
>>>>>>> +	xa_for_each(&pin->dpll_refs, i, ref) {
>>>>>>> +		const struct dpll_pin_ops *ops = dpll_pin_ops(ref);
>>>>>>> +		struct dpll_device *dpll = ref->dpll;
>>>>>>> +
>>>>>>> +		if (!ops->phase_adjust_set)
>>>>>>> +			return -EOPNOTSUPP;
>>>>>>
>>>>>>I'm thinking about this part. We can potentially have dpll devices with
>>>>>>different expectations on phase adjustments, right? And if one of them
>>>>>>won't be able to adjust phase (or will fail in the next line), then
>>>>>>netlink will return EOPNOTSUPP while _some_ of the devices will be
>>>>>>adjusted. Doesn't look great. Can we think about different way to apply
>>>>>>the change?
>>>>>>
>>>>>
>>>>>Well makes sense to me.
>>>>>
>>>>>Does following makes sense as a fix?
>>>>>We would call op for all devices which has been provided with the op.
>>>>>If device has no op -> add extack error, continue
>>>>
>>>>Is it real to expect some of the device support this and others don't?
>>>>Is it true for ice?
>>>>If not, I would got for all-or-nothing here.
>>>>
>>>
>>>Let's step back a bit.
>>>The op itself is introduced as per pin-dpll tuple.. did this
>>>intentionally,
>>>to inform each dpll that the offset has been changed - in case dplls are
>>>controlled by separated driver/firmware instances but still sharing the
>>>pin.
>>>Same way a pin frequency is being set, from user perspective on a pin, but
>>>callback is called for each dpll the pin was registered with.
>>>Whatever we do here, it shall be probably done for frequency_set()
>>>callback as
>>>well.
>>>
>>>The answers:
>>>So far I don't know the device that might do it this way, it rather
>>>supports
>>>phase_adjust or not. In theory we allow such behavior to be implemented,
>>>i.e.
>>>pin is registered with 2 dplls, one has the callback, second not.
>>
>>If there is only theoretical device like that now, implement
>>all-or-nothing. If such theoretical device appears in real, this could
>>be changed. The UAPI would not change, no problem.
>>
>
>I can live with it :)
>
>>
>>>Current hardware of ice sets phase offset for a pin no matter on which
>>>dpll
>>>device callback was invoked.
>>>"all-or-nothing" - do you mean to check all callback returns and then
>>>decide
>>>if it was successful?
>>
>>Check if all dplls have ops and only perform the action in such case. In
>>case one of the dplls does not have the op filled, return -EOPNOTSUPP.
>>
>>
>>Regarding the successful/failed op, I think you can just return. In
>>these cases, when user performs multiaction cmd, he should be prepared
>>to deal with consequences if part of this cmd fails. We don't have
>>rollback for any other multiaction cmd in dpll, I don't see why this
>>should be treated differently.
>>
>
>We don't have it because no one have spotted it on review,
>as mentioned the frequency_set behaves the same way,
>we need one approach for all of those cases.
>I am opting for having the rollback as suggested on the other thread.

Okay, but let's do that consistently.

>
>Thank you!
>Arkadiusz
>
>>
>>>
>>>Thank you!
>>>Arkadiusz
>>>
>>>>
>>>>>If device fails to set -> add extack error, continue
>>>>>Function always returns 0.
>>>>>
>>>>>Thank you!
>>>>>Arkadiusz
>>>>>
>>>>>>
>>>>>>> +		ret = ops->phase_adjust_set(pin,
>>>>>>> +					    dpll_pin_on_dpll_priv(dpll, pin),
>>>>>>> +					    dpll, dpll_priv(dpll), phase_adj,
>>>>>>> +					    extack);
>>>>>>> +		if (ret)
>>>>>>> +			return ret;
>>>>>>> +	}
>>>>>>> +	__dpll_pin_change_ntf(pin);
>>>>>>> +
>>>>>>> +	return 0;
>>>>>>> +}
>>>>>>> +
>>>