[PATCH 0/2] Remove input_call_hotplug

[PATCH 0/2] Remove input_call_hotplug

[Hannes Reinecke]

Hi all,

the input subsystem is using call_usermodehelper directly, which breaks 
all sorts of assertions especially when using udev.
And it's definitely going to fail once someone is trying to use netlink 
messages for hotplug event delivery.

To remedy this I've implemented a new sysfs class 'input_device' which 
is a representation of 'struct input_dev'. So each device listed in 
'/proc/bus/input/devices' gets a class device associated with it.
And we'll get proper hotplug events for each input_device which can be 
handled by udev accordingly.

Drawback is that a new event type (the said 'input_device') is added, so 
that hotplug scripts and udev might need to be adapted to handle it 
properly. And each device driver needs to be touched to write something 
meaningful as the class_id. A fallback is provided, but by neccessity is 
not very informative.

Patch 1/2 implements the core changes to drivers/input/input.c
Patch 2/2 provides proper device names for input drivers.

Patches are relative to bk://kernel.bkbits.net/vojtech/input

Comments are welcome.
Please CC me directly as I'm not on this list.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke			hare@suse.de
SuSE Linux AG				S390 & zSeries
Maxfeldstraße 5				+49 911 74053 688
90409 Nürnberg				http://www.suse.de

Re: [PATCH 0/2] Remove input_call_hotplug

[Greg KH]

On Tue, Jan 18, 2005 at 03:56:35PM +0100, Hannes Reinecke wrote:
> Hi all,
> 
> the input subsystem is using call_usermodehelper directly, which breaks 
> all sorts of assertions especially when using udev.
> And it's definitely going to fail once someone is trying to use netlink 
> messages for hotplug event delivery.
> 
> To remedy this I've implemented a new sysfs class 'input_device' which 
> is a representation of 'struct input_dev'. So each device listed in 
> '/proc/bus/input/devices' gets a class device associated with it.
> And we'll get proper hotplug events for each input_device which can be 
> handled by udev accordingly.

Hm, why another input class?  We already have /sys/class/input, which we
get hotplug events for.  We also have the individual input device
hotplug events, which is what I think we really want here, right?

Hm, in looking at the code some more, I think just adding a struct
device to the struct input_device would work.  You will have to refactor
out the fact that there is a pointer to a struct device in there (that
will just become the parent pointer to the input_dev's struct device).
If you do that, you can also drop a few of the fields in struct
input_dev.

That would solve all of the userspace issues your patch causes (hotplug
events would come out with the same type as before) and finally the
input subsystem would be part of the driver model properly :)

But yes, it will take more work than your patch, sorry.

thanks,

greg k-h

Re: [PATCH 0/2] Remove input_call_hotplug

[Dmitry Torokhov]

On Tue, 18 Jan 2005 13:30:02 -0800, Greg KH <greg@kroah.com> wrote:
> On Tue, Jan 18, 2005 at 03:56:35PM +0100, Hannes Reinecke wrote:
> > Hi all,
> >
> > the input subsystem is using call_usermodehelper directly, which breaks
> > all sorts of assertions especially when using udev.
> > And it's definitely going to fail once someone is trying to use netlink
> > messages for hotplug event delivery.
> >
> > To remedy this I've implemented a new sysfs class 'input_device' which
> > is a representation of 'struct input_dev'. So each device listed in
> > '/proc/bus/input/devices' gets a class device associated with it.
> > And we'll get proper hotplug events for each input_device which can be
> > handled by udev accordingly.
> 
> Hm, why another input class?  We already have /sys/class/input, which we
> get hotplug events for.  We also have the individual input device
> hotplug events, which is what I think we really want here, right?

These are a bit different classes. One is a generic input device class
device. Then you have several class device interfaces (evdev,
mousedev, joydev, tsdev, keyboard) that together with generic input
device produce concrete input devices (mouse, js, ts) that you have
implemented with class_simple.

At least that's the picture I had in my mind at some point.

-- 
Dmitry

Re: [PATCH 0/2] Remove input_call_hotplug

[Greg KH]

On Tue, Jan 18, 2005 at 04:49:34PM -0500, Dmitry Torokhov wrote:
> On Tue, 18 Jan 2005 13:30:02 -0800, Greg KH <greg@kroah.com> wrote:
> > On Tue, Jan 18, 2005 at 03:56:35PM +0100, Hannes Reinecke wrote:
> > > Hi all,
> > >
> > > the input subsystem is using call_usermodehelper directly, which breaks
> > > all sorts of assertions especially when using udev.
> > > And it's definitely going to fail once someone is trying to use netlink
> > > messages for hotplug event delivery.
> > >
> > > To remedy this I've implemented a new sysfs class 'input_device' which
> > > is a representation of 'struct input_dev'. So each device listed in
> > > '/proc/bus/input/devices' gets a class device associated with it.
> > > And we'll get proper hotplug events for each input_device which can be
> > > handled by udev accordingly.
> > 
> > Hm, why another input class?  We already have /sys/class/input, which we
> > get hotplug events for.  We also have the individual input device
> > hotplug events, which is what I think we really want here, right?
> 
> These are a bit different classes. One is a generic input device class
> device. Then you have several class device interfaces (evdev,
> mousedev, joydev, tsdev, keyboard) that together with generic input
> device produce concrete input devices (mouse, js, ts) that you have
> implemented with class_simple.

Hm, but we still need to make the input_dev a "real" struct device,
right?  And if you do that, then you just hooked up your hotplug event
properly, with no userspace breakage.

Then, if you want to still make the evdev, mousedev, and so on as
class_device interfaces, that's fine, but the main point of this patch
was to allow the call_usermodehelper call to be removed, so that the
input subsytem will work properly with the kernel event and hotplug
systems.

thanks,

greg k-h

Re: [PATCH 0/2] Remove input_call_hotplug

[Dmitry Torokhov]

On Tue, 18 Jan 2005 13:58:20 -0800, Greg KH <greg@kroah.com> wrote:
> On Tue, Jan 18, 2005 at 04:49:34PM -0500, Dmitry Torokhov wrote:
> > On Tue, 18 Jan 2005 13:30:02 -0800, Greg KH <greg@kroah.com> wrote:
> > > On Tue, Jan 18, 2005 at 03:56:35PM +0100, Hannes Reinecke wrote:
> > > > Hi all,
> > > >
> > > > the input subsystem is using call_usermodehelper directly, which breaks
> > > > all sorts of assertions especially when using udev.
> > > > And it's definitely going to fail once someone is trying to use netlink
> > > > messages for hotplug event delivery.
> > > >
> > > > To remedy this I've implemented a new sysfs class 'input_device' which
> > > > is a representation of 'struct input_dev'. So each device listed in
> > > > '/proc/bus/input/devices' gets a class device associated with it.
> > > > And we'll get proper hotplug events for each input_device which can be
> > > > handled by udev accordingly.
> > >
> > > Hm, why another input class?  We already have /sys/class/input, which we
> > > get hotplug events for.  We also have the individual input device
> > > hotplug events, which is what I think we really want here, right?
> >
> > These are a bit different classes. One is a generic input device class
> > device. Then you have several class device interfaces (evdev,
> > mousedev, joydev, tsdev, keyboard) that together with generic input
> > device produce concrete input devices (mouse, js, ts) that you have
> > implemented with class_simple.
> 
> Hm, but we still need to make the input_dev a "real" struct device,
> right?  And if you do that, then you just hooked up your hotplug event
> properly, with no userspace breakage.

I wasn't planning on doing that. The real devices are serio ports,
gameport ports and USB devices.They require power and resource
management and so forth. input_device is just a product of binding a
port to appropriate driver and seems to me like an ideal class_device
candidate. Then you add couple of class interfaces and get another
class_device layer as a result.

> Then, if you want to still make the evdev, mousedev, and so on as
> class_device interfaces, that's fine, but the main point of this patch
> was to allow the call_usermodehelper call to be removed, so that the
> input subsytem will work properly with the kernel event and hotplug
> systems.
>

I was mostly talking about the need of 2 separate classes and this
patch lays groundwork for it althou lifetime rules in input system
need to be cleaned up before we can go all the way.

-- 
Dmitry

Re: [PATCH 0/2] Remove input_call_hotplug

[Greg KH]

On Tue, Jan 18, 2005 at 05:20:40PM -0500, Dmitry Torokhov wrote:
> On Tue, 18 Jan 2005 13:58:20 -0800, Greg KH <greg@kroah.com> wrote:
> > On Tue, Jan 18, 2005 at 04:49:34PM -0500, Dmitry Torokhov wrote:
> > > On Tue, 18 Jan 2005 13:30:02 -0800, Greg KH <greg@kroah.com> wrote:
> > > > On Tue, Jan 18, 2005 at 03:56:35PM +0100, Hannes Reinecke wrote:
> > > > > Hi all,
> > > > >
> > > > > the input subsystem is using call_usermodehelper directly, which breaks
> > > > > all sorts of assertions especially when using udev.
> > > > > And it's definitely going to fail once someone is trying to use netlink
> > > > > messages for hotplug event delivery.
> > > > >
> > > > > To remedy this I've implemented a new sysfs class 'input_device' which
> > > > > is a representation of 'struct input_dev'. So each device listed in
> > > > > '/proc/bus/input/devices' gets a class device associated with it.
> > > > > And we'll get proper hotplug events for each input_device which can be
> > > > > handled by udev accordingly.
> > > >
> > > > Hm, why another input class?  We already have /sys/class/input, which we
> > > > get hotplug events for.  We also have the individual input device
> > > > hotplug events, which is what I think we really want here, right?
> > >
> > > These are a bit different classes. One is a generic input device class
> > > device. Then you have several class device interfaces (evdev,
> > > mousedev, joydev, tsdev, keyboard) that together with generic input
> > > device produce concrete input devices (mouse, js, ts) that you have
> > > implemented with class_simple.
> > 
> > Hm, but we still need to make the input_dev a "real" struct device,
> > right?  And if you do that, then you just hooked up your hotplug event
> > properly, with no userspace breakage.
> 
> I wasn't planning on doing that. The real devices are serio ports,
> gameport ports and USB devices.They require power and resource
> management and so forth. input_device is just a product of binding a
> port to appropriate driver and seems to me like an ideal class_device
> candidate. Then you add couple of class interfaces and get another
> class_device layer as a result.

Ah, ok, that makes sense.  That would work too, although I don't know if
udev can handle class_interfaces with a "dev" file in it or not.  If
not, it shouldn't be that hard to change.

> > Then, if you want to still make the evdev, mousedev, and so on as
> > class_device interfaces, that's fine, but the main point of this patch
> > was to allow the call_usermodehelper call to be removed, so that the
> > input subsytem will work properly with the kernel event and hotplug
> > systems.
> >
> 
> I was mostly talking about the need of 2 separate classes and this
> patch lays groundwork for it althou lifetime rules in input system
> need to be cleaned up before we can go all the way.

I agree.  But I think only 1 class is needed, that way we don't break
userspace, which is a pretty important thing.

thanks,

greg k-h

Re: [PATCH 0/2] Remove input_call_hotplug

[Hannes Reinecke]

Greg KH wrote:
> On Tue, Jan 18, 2005 at 05:20:40PM -0500, Dmitry Torokhov wrote:
> 
>>On Tue, 18 Jan 2005 13:58:20 -0800, Greg KH <greg@kroah.com> wrote:
>>
>>>On Tue, Jan 18, 2005 at 04:49:34PM -0500, Dmitry Torokhov wrote:
>>>
>>>>On Tue, 18 Jan 2005 13:30:02 -0800, Greg KH <greg@kroah.com> wrote:
>>>>
>>>>>On Tue, Jan 18, 2005 at 03:56:35PM +0100, Hannes Reinecke wrote:
>>>>>
>>>>>>Hi all,
>>>>>>
>>>>>>the input subsystem is using call_usermodehelper directly, which breaks
>>>>>>all sorts of assertions especially when using udev.
>>>>>>And it's definitely going to fail once someone is trying to use netlink
>>>>>>messages for hotplug event delivery.
>>>>>>
>>>>>>To remedy this I've implemented a new sysfs class 'input_device' which
>>>>>>is a representation of 'struct input_dev'. So each device listed in
>>>>>>'/proc/bus/input/devices' gets a class device associated with it.
>>>>>>And we'll get proper hotplug events for each input_device which can be
>>>>>>handled by udev accordingly.
>>>>>
>>>>>Hm, why another input class?  We already have /sys/class/input, which we
>>>>>get hotplug events for.  We also have the individual input device
>>>>>hotplug events, which is what I think we really want here, right?
>>>>
>>>>These are a bit different classes. One is a generic input device class
>>>>device. Then you have several class device interfaces (evdev,
>>>>mousedev, joydev, tsdev, keyboard) that together with generic input
>>>>device produce concrete input devices (mouse, js, ts) that you have
>>>>implemented with class_simple.
>>>
>>>Hm, but we still need to make the input_dev a "real" struct device,
>>>right?  And if you do that, then you just hooked up your hotplug event
>>>properly, with no userspace breakage.
>>
>>I wasn't planning on doing that. The real devices are serio ports,
>>gameport ports and USB devices.They require power and resource
>>management and so forth. input_device is just a product of binding a
>>port to appropriate driver and seems to me like an ideal class_device
>>candidate. Then you add couple of class interfaces and get another
>>class_device layer as a result.
> 
> 
> Ah, ok, that makes sense.  That would work too, although I don't know if
> udev can handle class_interfaces with a "dev" file in it or not.  If
> not, it shouldn't be that hard to change.
> 
> 
>>>Then, if you want to still make the evdev, mousedev, and so on as
>>>class_device interfaces, that's fine, but the main point of this patch
>>>was to allow the call_usermodehelper call to be removed, so that the
>>>input subsytem will work properly with the kernel event and hotplug
>>>systems.
>>>
>>
>>I was mostly talking about the need of 2 separate classes and this
>>patch lays groundwork for it althou lifetime rules in input system
>>need to be cleaned up before we can go all the way.
> 
> 
> I agree.  But I think only 1 class is needed, that way we don't break
> userspace, which is a pretty important thing.
> 
Well, if you could show me how to do this with the class_interface thing 
I'd be happy to comply.

The input layer design is like this:
- Physical devices present one (or several) abstract input devices 
(which correspond to struct input_dev)
- Each input device can be linked to one or several input handlers
(which correspond to struct input_handle)
- Each handler is represented to userspace with a device node.

The problem with the current input layer is that each 'struct 
input_handle' is associated with a class_simple device.
This class is named 'input', so we're getting 'input' events from it.
But each instantiation of struct input_dev is also sending 'input' 
events as it is doing the call_usermodehelper call directly.

Effectively we're having _two_ different hotplug events with the same 
name, which is inconsistent in itself. I don't mind at all if we're 
breaking this.

But if we were going to implement this with device_interface, we'd be 
having a /sys/class structure like:

class
|- isa0060-serio0-input0
|  |- event0
|  |  `dev
|  |- key
|  |- ...
|- ..

So we'd be moving the 'dev' attribute one directory down, again 
incurring a userland breakage.
Plus it would be far more coding involved as the entire input layer 
structure would have to be redone.

But if the powers that be decide for the latter option, who am I to 
argue ...
Only I'd like to have this resolved quite soon as it effectively blocks 
from switching to netlink messages for hotplug events completely.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke			hare@suse.de
SuSE Linux AG				S390 & zSeries
Maxfeldstraße 5				+49 911 74053 688
90409 Nürnberg				http://www.suse.de

Re: [PATCH 0/2] Remove input_call_hotplug

[Dmitry Torokhov]

On Wed, 19 Jan 2005 12:56:16 +0100, Hannes Reinecke <hare@suse.de> wrote:
> Greg KH wrote:
> > On Tue, Jan 18, 2005 at 05:20:40PM -0500, Dmitry Torokhov wrote:
> >
> >>I was mostly talking about the need of 2 separate classes and this
> >>patch lays groundwork for it althou lifetime rules in input system
> >>need to be cleaned up before we can go all the way.
> >
> >
> > I agree.  But I think only 1 class is needed, that way we don't break
> > userspace, which is a pretty important thing.
> > 
> Well, if you could show me how to do this with the class_interface thing
> I'd be happy to comply.
> 
> The input layer design is like this:
> - Physical devices present one (or several) abstract input devices
> (which correspond to struct input_dev)
> - Each input device can be linked to one or several input handlers
> (which correspond to struct input_handle)
> - Each handler is represented to userspace with a device node.
> 
> The problem with the current input layer is that each 'struct
> input_handle' is associated with a class_simple device.
> This class is named 'input', so we're getting 'input' events from it.
> But each instantiation of struct input_dev is also sending 'input'
> events as it is doing the call_usermodehelper call directly.
> 

Yes, this is the problem and needs to be resolved. Thankfully most
people have keyboard support compiled in so it's not fatal but we
probably waht hotplug package be updated first before we commit this
change.

> 
> But if we were going to implement this with device_interface, we'd be
> having a /sys/class structure like:
> 
> class
> |- isa0060-serio0-input0
> |  |- event0
> |  |  `dev
> |  |- key
> |  |- ...
> |- ..
> 
> So we'd be moving the 'dev' attribute one directory down, again
> incurring a userland breakage.
> Plus it would be far more coding involved as the entire input layer
> structure would have to be redone.
> 

If I understand correctly we do not have subclasses so it will look like
class
|- input_device
|  |- input0
|  |- input1
|
|- input
|  |-event0
|  |-event1
|  |-mouse0

So breakage is really minimal.

-- 
Dmitry

Re: [PATCH 0/2] Remove input_call_hotplug

[Greg KH]

On Wed, Jan 19, 2005 at 09:30:14AM -0500, Dmitry Torokhov wrote:
> If I understand correctly we do not have subclasses so it will look like
> class
> |- input_device
> |  |- input0
> |  |- input1
> |
> |- input
> |  |-event0
> |  |-event1
> |  |-mouse0
> 
> So breakage is really minimal.

I really want classes to be able to have "parent" classes someday.  It's
not that tough of a change to the driver core, if someone wants to do
the work.  That would enable this input stuff to be cleaner, right?

It would also allow us to move the /sys/block stuff to use classes,
which it can't right now due to the lack of the parent ability.

thanks,

greg k-h