[v2,2/2] PCI / ACPI: PCI delay optimization from ACPI

On 03/24/2015 10:08 PM, Bjorn Helgaas wrote:
> On Tue, Mar 24, 2015 at 05:04:58PM +0800, Aaron Lu wrote:
>> diff --git a/drivers/pci/pci-acpi.c b/drivers/pci/pci-acpi.c
>> index e0afc94aca01..220371c2def4 100644
>> --- a/drivers/pci/pci-acpi.c
>> +++ b/drivers/pci/pci-acpi.c
>> @@ -537,11 +537,24 @@ static struct pci_platform_pm_ops acpi_pci_platform_pm = {
>>  
>>  void acpi_pci_add_bus(struct pci_bus *bus)
>>  {
>> +	union acpi_object *obj;
>> +
>>  	if (acpi_pci_disabled || !bus->bridge)
>>  		return;
>>  
>>  	acpi_pci_slot_enumerate(bus);
>>  	acpiphp_enumerate_slots(bus);
>> +
>> +	/*
>> +	 * For a host bridge, check its _DSM for function 8 and if
>> +	 * that is available, mark it in the corresponding pci_bus.
>> +	 */
>> +	if (bus->bridge->parent)
>> +		return;
> 
> This is not really an obvious way of testing for a host bridge.  I think
> pci_is_root_bus() would be a better way, but I'm still hoping for something
> in pci_root.c instead.  There is find_pci_host_bridge(), which might be
> useful (it's currently static but we might want to rename and export it for
> this and other reasons).

I agree that pci_root.c is the proper place for a property of the PCI host
bridge, but since it only calls:
	root->bus = pci_acpi_scan_root(root);
and when it returns, eveything is already properly initialized. To pass
the ignore_reset_delay information from there, we probably need to add a
field ignore_reset_delay to struct acpi_pci_root and then pass that
information to pci_root_info and then pass the pointer of pci_root_info
to pci_create_root_bus where we use that information to set the field in
the pci_host_bridge. Intead of doing it this way, hook the
acpi_pci_add_bus to set the ignore_reset_delay bit of the PCI host
bridge seems to be cheap and easier.

> 
>> +	obj = acpi_evaluate_dsm(ACPI_HANDLE(bus->bridge), pci_acpi_dsm_uuid, 3,
>> +				RESET_DELAY_DSM, NULL);
>> +	if (obj && obj->type == ACPI_TYPE_INTEGER && obj->integer.value == 1)
>> +		bus->ignore_reset_delay = 1;
> 
> I think you need to free "obj" here.  Other acpi_evaluate_dsm() callers use
> ACPI_FREE().

Oops yes.

> 
>>  }
>>  
>>  void acpi_pci_remove_bus(struct pci_bus *bus)
>> @@ -567,6 +580,55 @@ static struct acpi_device *acpi_pci_find_companion(struct device *dev)
>>  				      check_children);
>>  }
>>  
>> +/**
>> + * pci_acpi_delay_optimize - optimize PCI D3 and D3cold delay from ACPI
>> + * @pdev: the PCI device whose delay is to be updated
>> + * @adev: the companion ACPI device of this PCI device
>> + *
>> + * Update the d3_delay and d3cold_delay of a PCI device from the ACPI _DSM
>> + * control method of either its own or its parent bridge.
>> + *
>> + * The UUID of the _DSM control method, together with other information like
>> + * which delay values can be optimized, etc. is defined in a ECN available on
>> + * PCIsig.com titled as: ACPI additions for FW latency optimizations.
>> + * Function 9 of the ACPI _DSM control method, if available for a specific PCI
>> + * device, provides various possible delay values that are less than what the
>> + * SPEC requires. Here, we only deal with d3_delay and d3cold_delay. Others
>> + * can be added later.
>> + * Function 8 of the ACPI _DSM control method, if available for the PCI host
>> + * bridge(reflected by the bus' ignore_reset_delay filed), means all its
>> + * children devices do not need the reset delay when leaving from D3cold state.
>> + */
>> +static void pci_acpi_delay_optimize(struct pci_dev *pdev,
>> +				    acpi_handle handle)
>> +{
>> +	int value;
>> +	union acpi_object *obj, *elements;
>> +
>> +	if (pdev->bus->ignore_reset_delay)
>> +		pdev->d3cold_delay = 0;
> 
> I think this only propagates the function 8 result to the immediate
> children of the host bridge, i.e., devices on the root bus.  But the ECN
> says it affects the entire hierarchy.  Can you put the ignore_reset_delay
> bit in the struct pci_host_bridge instead?

No problem.

> 
>> +
>> +	obj = acpi_evaluate_dsm(handle, pci_acpi_dsm_uuid, 3,
>> +				FUNCTION_DELAY_DSM, NULL);
>> +	if (!obj)
>> +		return;
>> +
>> +	if (obj->type == ACPI_TYPE_PACKAGE && obj->package.count == 5) {
>> +		elements = obj->package.elements;
>> +		if (elements[0].type == ACPI_TYPE_INTEGER) {
>> +			value = (int)elements[0].integer.value / 1000;
>> +			if (value < PCI_PM_D3COLD_WAIT)
>> +				pdev->d3cold_delay = value;
>> +		}
>> +		if (elements[3].type == ACPI_TYPE_INTEGER) {
>> +			value = (int)elements[3].integer.value / 1000;
>> +			if (value < PCI_PM_D3_WAIT)
>> +				pdev->d3_delay = value;
>> +		}
>> +	}
>> +	kfree(obj);
>> +}
>> +
>>  static void pci_acpi_setup(struct device *dev)
>>  {
>>  	struct pci_dev *pci_dev = to_pci_dev(dev);
>> @@ -575,6 +637,9 @@ static void pci_acpi_setup(struct device *dev)
>>  	if (!adev)
>>  		return;
>>  
>> +	if (pci_dev->pm_cap)
>> +		pci_acpi_delay_optimize(pci_dev, adev->handle);
> 
> Is the "pm_cap" test really necessary?  If we do it this way, we then have
> to convince ourselves that pdev->d3cold_delay and pdev->d3_delay are only
> needed when pdev has a pm_cap.
> 
> If we *always* fill in the delay values, it's possible they won't be used,
> but we don't have to prove any connection between them and a pm_cap, so
> the code is easier to analyze.

Brilliant.

Here is an updat, kind of a proof of concept one, to see if it goes the
right way:

[v2,2/2] PCI / ACPI: PCI delay optimization from ACPI

Commit Message

Comments

Patch