[systemd-devel] [WIP PATCH 2/4] ACPI: button: remove the LID input node when the state is unknown

[Zheng, Lv]

Hi, Benjamin

> From: linux-acpi-owner at vger.kernel.org [mailto:linux-acpi-owner at vger.kernel.org] On Behalf Of Benjamin
> Tissoires
> Subject: Re: [WIP PATCH 2/4] ACPI: button: remove the LID input node when the state is unknown
> 
> Hi Lv,
> 
> On Jun 05 2017 or thereabouts, Zheng, Lv wrote:
> > Hi, Benjamin
> >
> > > From: Benjamin Tissoires [mailto:benjamin.tissoires at redhat.com]
> > > Subject: [WIP PATCH 2/4] ACPI: button: remove the LID input node when the state is unknown
> 
> > My dell latitude 6430u test platform sends multiple Notify(lid) before suspend and after resume.
> 
> Does this platform requires the not lid_reliable check as per this
> series? Because if it doesn't, then we should not care.

No need to mark lid_reliable.

> > This is because the aml table puts many Notify(LID, 0x80) in various control methods.
> > And not one of them but multiple of them will be invoked by various OS drivers during suspend/resume
> period.
> > I think this is not an isolated platform that will invoke multiple redundant "Notify(lid)".
> >
> > Fortunately, the lid state for the multiple notify(lid) should be same as the first "Notify(lid)".
> > I suppose this is why SW_LID is invented, as it can really filter such redundant events.
> > And user space finally can only see 1 "close" event.
> >
> > But unconditionally prepending "open" before all "close" events surely can break the logic by
> > delivering multiple "close" events to the user space.
> 
> That doesn't matter. What matters is the state of the switch, not the
> event. So if user space receives (in case we marked the switch as not
> reliable) several close events, all user space will do is realize that
> the state is still closed and will act accordingly.

OK, I tried to address this here:
https://patchwork.kernel.org/patch/9771121/

> > Another issue is, for case 5, when we use button.lid_init_state=method.
> > Unconditionally prepending "open" before driver initiated "close" event
> > sent due to acpi_lid_initialize_state(), we will see suspend/resume cycles.
> 
> Case 5 is broken anyway and needs to be handled specially. It was not
> targeted in this WIP series.

It was addressed by button.lid_init_state=open and newer systemd.
It's not broken any more.


> > Thus if we consider both cases, we should:
> > 1. put a frequency check to filter possible redundant events.
> 
> This doesn't work and should be avoided. The state of the input switch
> is known to the input layer only, and given there are spinlocks, you can
> not know if the state is actually the one you expected beforehand.
> 
> You can however add frequency checks in the input handler, but that
> would assume the input layer is not doing its job properly and so should
> be avoided.

OK, I dropped frequency check mechanism.

> > 2. distinguish AML "Notify" call and button driver initiated lid notification.
> 
> Again, we don't care if the "event" comes from ACPI, the driver itself or
> user space (libinput). All that matters is the current state of the
> input node switch, that needs to match the physical world at any time.

That depends on the final test result.
However I managed to get systemd working with case 2,4 using this commit:
https://patchwork.kernel.org/patch/9771121/

> > This is another major differences between your proposal and mine.
> >
> > First of all, I think it should be in a separate patch.
> 
> Well, that's already a patch on its own :/
> 
> >
> > Second, I have concerns related to such a change:
> > I can see that, you are trying to address a problem that:
> > The input layer requires a determined initial SW_LID state while ACPI button driver cannot offer.
> > So by adding/removing input node, you can introduce a tristate SW_LID input node.
> 
> You can put it that way. I prefer putting it: "when we export the LID
> switch input node, you are guaranteed to have the proper state".
> 
> > However I doubt if this is necessary and can solve real issues, as:
> > systemd now works fine with button driver for all cases,
> 
> I do not care about systemd or the suspend lopps introduced by systemd.
> All I care is that the kernel provides correct behavior. If systemd can
> work around some issues we see because we are too lazy to fix them in
> the kernel (this is not a personal attack, sometimes being lazy is the
> right solution), fine. But the current state of this driver doesn't
> follow the specification of the input subsystem on some platform, and
> this is what this series fixes.

No, we just can see if case 5 is properly addressed (like current systemd 233 does),
we don't need to do anything.
So if you still insist to fix systemd 229, I just post an RFC:
https://patchwork.kernel.org/patch/9771121/

> > only desktop managers should be changed to be compliant to case 2/4/5
> 
> As long as the kernel lies, we should not even remotely envision asking
> user space to change.

By reverting back to "method" mode and fixing "method" mode, there is no regression.
What if we just stay to see what will happen next with MS Surface Pro 1 docking station support.
Maybe it is minor.

> > (or if we improve by adding a timer, they should only be changed to be compliant to case 5).
> > And doing this won't help desktop managers to be able to work fine with case 5.
> > So this finally may only remove ACPI SW_LID support from Surface Pro 1.
> 
> I haven't decided how to solve case 5 completely. So please do not take
> this case into account yet.
> 
> > While with latest systemd, closing lid on that platform can correctly triggers suspend.
> > And no suspend/resume cycles should be seen.
> > So why do we need to remove this feature (ACPI SW_LID) from Surface Pro 1?
> 
> Well, with this change, if you mark the surface pro 1 as unreliable, the
> event will be forwarded to user space correctly. Just that there is a
> systemd bug in which the state is not synced when the device appears.
> 
> So yes, it'll expose a new bug in userspace, but given the blacklist of
> unreliable LID switches will be handled in userspace, user space can
> make sure this will not show up before systemd can handle it.

IMO, we don't need to remove input node, button.lid_init_state=ignore is developed for the same purpose.
And it works perfectly with systemd 233. So we can stay with old solution.
So hope this can replace your aggressive change:
https://patchwork.kernel.org/patch/9771119/

> > If you really want to propose an ABI change for user space.
> > Why don't you do this in input layer by defining SW_LID as tristate?
> 
> Because this proposal is not a kABI change, and the kABI change you
> propose is just not doable. We have too many users of EV_SW to be able
> to say that we change the semantic. A solution would be to add a new
> EV_* event, but we don't really need.

Cheers,
Lv