On 01-07-20, 10:25, Pierre-Louis Bossart wrote:
* wake up master and slave so that slave can notify master* the wakeen event and let codec driver check codec status*/- list_for_each_entry(slave, &bus->slaves, node) {
/** discard devices that are defined in ACPI tables but* not physically present and devices that cannot* generate wakes*/if (slave->dev_num_sticky && slave->prop.wake_capable)pm_request_resume(&slave->dev);Hmmm, shouldn't slave do this? would it not make sense to notify the slave thru callback and then slave decides to resume or not..?
In this mode, the bus is clock-stop mode, and events are detected with level detector tied to PCI events. The master and slave devices are all in pm_runtime suspended states. The codec cannot make any decisions on its own since the bus is stopped, it needs to first resume, which assumes that the master resumes first and the enumeration re-done before it can access any of its registers.
By looping through the list of devices that can generate events, you end-up first forcing the master to resume, and then each slave resumes and can check who generated the event and what happened while suspended. if the codec didn't generate the event it will go back to suspended mode after the usual timeout.
We can add a callback but that callback would only be used for Intel solutions, but internally it would only do a pm_request_resume() since the codec cannot make any decisions before first resuming. In other words, it would be an Intel-specific callback that is implemented using generic resume operations. It's probably better to keep this in Intel-specific code, no?
I do not like the idea that a device would be woken up, that kind of defeats the whole idea behind the runtime pm. Waking up a device to check the events is a generic sdw concept, I don't see that as Intel specific one.
In this case, this in an Intel-specific mode that's beyond what MIPI defined. This is not the traditional in-band SoundWire wake defined in the SoundWire specification. The bus is completely down, the master IP is powered-off and all context lost. There is still the ability for a Slave device to throw a wake as defined by MIPI in clock-stop-mode1, but this is handled with a separate level detector and the wake detection is handled by the PCI subsystem. On a wake, the master IP needs to be powered-up, the entire bus needs to be restarted and the Slave devices re-enumerated.
Right and I would expect that Slave device would do runtime_get_sync() first thing in the callback. That would ensure that the Master is powered up, Slave is powered up, all enumeration is complete. This is more standard way to deal with this, we expect devices to be so we low powered or off so cannot do read/write unless we resume.
We trigger that sequence with a loop that calls pm_request_resume() for all Slave devices that are present and exposed in their properties that they are wake-capable. As you rightly said in your comments, this will result a nice wake-up handled by the pm_runtime framework in the right sequence (DSP first, then SoundWire master then Slave devices).
You will see in follow-up patches that the master IP can be configured in different clock stop modes, depending on the needs (power/latency compromise typically). When the more traditional SoundWire wake is used, we do not use this sequence at all.
The point which is not clear to me if why do we need a specific sequence. Above sequence should work for you, if not I would like to understand why.
I would like to see a generic callback for the devices and let devices do the resume part, that is standard operating principle when we deal with suspended devices. If the device thinks they need to resume, they will do the runtime resume, check the status and sleep if not applicable. Since we have set the parents correctly, any resume operation for slaves would wake master up as well...
I do not see a need for intel driver to resume slave devices here, or did I miss something?
Yes, the part "If the device thinks they need to resume" is not quite right. The Slave device cannot access its registers before fully resuming, which in this case means a re-enumeration, so cannot "think" or make a decision on its own. That's the reason why we force them to resume, since it's the first step they would need to do anyways, and then if they don't have anything to do they go back to idle after a timeout. I invite you to see the suspend/resume sequences in codec drivers where you will see regmap access moves to cache-only on suspend and full access restored on resume, along with a hardware sync.
I see your point and I really think we are talking about a similar sequence, but we simplified your idea since there's no possibility of making a decision before Slave devices resume.
If we are worried about Slave device remembering then we should save state in device context. But i think that slave should *always* perform runtime resume as a first step in the callback before they would do any bus/device ops. After all, the callback is wake from low power state
The only optimization we did here is that we only resume Slave devices than can generate a wake, and filter out the 'ghost' devices that are described in ACPI tables but don't physically exist.
Does this help?