Re: How to enable auto-suspend by default

[Mathias Nyman]

On 23.11.2020 15.54, Hans de Goede wrote:
> Hi,
> 
> On 11/11/20 3:31 PM, Mika Westerberg wrote:
>> On Wed, Nov 11, 2020 at 12:27:32PM +0100, Hans de Goede wrote:
>>> Hi,
>>>
>>> On 11/10/20 6:25 PM, Mika Westerberg wrote:
>>>> On Tue, Nov 10, 2020 at 04:02:33PM +0000, Limonciello, Mario wrote:
>>>>>>
>>>>>> On Tue, Nov 10, 2020 at 11:57:07AM +0100, Bastien Nocera wrote:
>>>>>>> Hey,
>>>>>>>
>>>>>>> systemd has been shipping this script to enable auto-suspend on a
>>>>>>> number of USB and PCI devices:
>>>>>>>
>>>>>> https://github.com/systemd/systemd/blob/master/tools/chromiumos/gen_autosuspen
>>>>>> d_rules.py
>>>>>>>
>>>>>>> The problem here is twofold. First, the list of devices is updated from
>>>>>>> ChromeOS, and the original list obviously won't be updated by ChromeOS
>>>>>>> developers unless a device listed exists in a ChromeBook computer,
>>>>>>> which means a number of devices that do support autosuspend aren't
>>>>>>> listed.
>>>>>>>
>>>>>>> The other problem is that this list needs to exist at all, and that it
>>>>>>> doesn't seem possible for device driver developers (at various levels
>>>>>>> of the stack) to opt-in to auto-suspend when all the variants of the
>>>>>>> device (or at least detectable ones) support auto-suspend.
>>>>>>
>>>>>> A driver can say they support autosuspend today, but I think you are
>>>>>> concerned about the devices that are controlled by class-compliant
>>>>>> drivers, right?  And for those, no, we can't do this in the kernel as
>>>>>> there are just too many broken devices out there.
>>>>>>
>>>>>
>>>>> I guess what Bastien is getting at is for newer devices supported by class
>>>>> drivers rather than having to store an allowlist in udev rules, can we set
>>>>> the allowlist in the kernel instead.  Then distributions that either don't
>>>>> use systemd or don't regularly update udev rules from systemd can take
>>>>> advantage of better defaults on modern hardware.
>>>>>
>>>>> The one item that stood out to me in that rules file was 8086:a0ed.
>>>>> It's listed as "Volteer XHCI", but that same device ID is actually present
>>>>> in an XPS 9310 in front of me as well and used by the xhci-pci kernel module.
>>>>>
>>>>> Given we're effectively ending up with the combination of runtime PM turned
>>>>> on by udev rules, do we need something like this for that ID:
>>>>>
>>>>> https://github.com/torvalds/linux/commit/6a7c533d4a1854f54901a065d8c672e890400d8a
>>>>>
>>>>> @Mika Westerberg should 8086:a0ed be quirked like the TCSS xHCI too?
>>>>
>>>> I think this one is the TGL PCH xHCI. The quirk currently for xHCI
>>>> controllers that are part of the TCSS (Type-C SubSystem) where it is
>>>> important to put all devices into low power mode whenever possible,
>>>> otherwise it keeps the whole block on.
>>>
>>> Note that there are currently some IDs missing from the xHCIs which
>>> are part of the TCSS too. At least the id for the xHCI in the thunderbolt
>>> controller on the Lenovo T14 gen 1 is missing. I started a discussion
>>> about extending the kernel quirk list for this vs switching to hwdb
>>> a while a go:
>>>
>>> https://lore.kernel.org/linux-usb/b8b21ba3-0a8a-ff54-5e12-cf8960651086@xxxxxxxxxx/
>>>
>>> The conclusion back then was to switch to hwdb, but I never got around to this.
>>
>> The reason I've added these to the xHCI driver is that it works even if
>> you are running some really small userspace (like busybox). Also for the
>> xHCI in TCSS we know for sure that it fully supports D3cold.
>>
>> (The one you refer above is actually mistake from my side as I never
>>  tested Alpine Ridge LP controller which I think this is).
> 
> Ok, so I'll submit a patch adding the 15c1 product-id for the
> INTEL_ALPINE_RIDGE_LP_2C_XHCI controller to the list of ids for which we
> set the XHCI_DEFAULT_PM_RUNTIME_ALLOW quirk. To fix the much too high
> idle-power consumption problem on devices with this Alpine Ridge variant.

Thanks

> 
>>>> Typically we haven't done that for PCH side xHCI controllers though, but
>>>> I don't see why not if it works that is. Adding Mathias to comment more
>>>> on that since he is the xHCI maintainer.
>>>
>>> If we are also going to enable this for the non TCSS Intel XHCI controllers,
>>> maybe just uncondtionally enable it for all Intel XHCI controllers, or
>>> if necessary do a deny-list for some older models and enable it for anything
>>> not on the deny-list (so all newer models). That should avoid the game of
>>> whack-a-mole which we will have with this otherwise.
>>
>> This is really up to Mathias to decide. I'm fine either way :)
> 
> Ok, Matthias what do you think about this?

I don't think we are ready to enable runtime pm as default for all Intel xHCI controllers.
The risk of xHCI not waking up when user plugs a mouse/keyboard, making the system unusable
just seems too high compared to the powersaving benefit.

The powersaving benefit from autosuspending the TCSS xHCI is a lot better, and we, (Mika mostly)
has been able to verify they work.

So I propose we for now continue adding TCSS xHCI controllers to the allowlist in kernel.
For others I think a userspace allow/denylist makes sense.

Long term goal would be default allow for all, with short denylist in kernel.

Thanks
Mathias