Re: 2.6.36/2.6.37: broken compatibility with userspace input-utils ?

From: Dmitry Torokhov
Date: Fri Jan 28 2011 - 13:34:21 EST


On Fri, Jan 28, 2011 at 04:15:51PM -0200, Mauro Carvalho Chehab wrote:
> Em 28-01-2011 15:33, Dmitry Torokhov escreveu:
> > On Fri, Jan 28, 2011 at 03:01:58PM -0200, Mauro Carvalho Chehab wrote:
> >> Em 28-01-2011 14:40, Dmitry Torokhov escreveu:
> >>> On Fri, Jan 28, 2011 at 09:55:58AM -0200, Mauro Carvalho Chehab wrote:
> >>
> >>>> The rc-core register (and the corresponding input register) is done when
> >>>> the device detected a remote controller, so, it should be safe to register
> >>>> on that point. If not, IMHO, there's a bug somewhere.
> >>>
> >>> It is not a matter of safe or unsafe registration. Registration is fine.
> >>> The problem is that with the current set up is that utility is fired
> >>> when trunk of [sub]tree is created, but the utility wants to operate on
> >>> leaves which may not be there yet.
> >>
> >> I'm not an udev expert. Is there a udev event that hits only after having
> >> the driver completely loaded?
> >
> > Define completely loaded? For a PCI SCSI controller does fully loaded
> > mean all attached devices are discovered and registered with block layer?
> > For a wireless NIC does it mean that it assocuated with an AP? What if
> > you have more than one device that driver serves?
> >
> > So teh answer is no and there should not be.
> >
> >>
> >> Starting an udev rule while modprobe is
> >> still running is asking for race conditions.
> >
> > Not if we write stuff properly.
> >
> >>
> >> I'm not entirely convinced that this is the bug that Mark is hitting, as
> >
> > I do not know yet.
> >
> >> rc-core does all needed setups before registering the evdev device. We
> >> need the core and the dmesg to be sure about what's happening there.
> >
> > I will say it again. Your udev rule triggers when you create rcX device.
> > eventX device may apeear 2 hours after that (I could have evdev as a
> > module and blacklisted and load it later manually).
>
> Blacklisting it won't (or shouldn't work).
>
> From rc-main, the registering sequence is:
>
> dev_set_name(&dev->dev, "rc%ld", dev->devno);
> dev_set_drvdata(&dev->dev, dev);
> rc = device_add(&dev->dev);
> if (rc)
> return rc;
>
> rc = ir_setkeytable(dev, rc_map);
> if (rc)
> goto out_dev;
>
> dev->input_dev->dev.parent = &dev->dev;
> memcpy(&dev->input_dev->id, &dev->input_id, sizeof(dev->input_id));
> dev->input_dev->phys = dev->input_phys;
> dev->input_dev->name = dev->input_name;
> rc = input_register_device(dev->input_dev);
> if (rc)
> goto out_table;
>
> rc-main will wait for input_register_device() to finish, so even if you
> blacklist it, rc-core will load it, in order to solve the symbol dependency.

No, input_register_device() and input core itself does not have symbol
dependency on evdev (which provides /dev/input/eventX), which is just an
input handler. A very important, but still an optional, handler.

> Btw, there's really a race issue there: device_add is happening before
> input_register_device(), so the udev rule will cause troubles.

That's what I have been saying in the last 3+ emails, yes.

>
> > You need to split it into 2 separate steps:
> >
> > 1. Triggers when rcX appears, accesses only rcX and it's parents and
> > does rcX related stuff.
> >
> > 2. Triggers when eventX appears and loads keymap and what not. Because
> > it is a child of rcX (in specific case of remotes) it may examine
> > rcX attributes as well.
>
> The fix is probably simpler: we need to change the udev rules to work at
> evdev registration and only if the device is a remote controller. This
> should solve the current issue.

The problem I have with it is that it violates layering. You affect
different subsystems/layers, why do you insist on jamming them togetehr?
Don't do the kitchen sink style, pretty please.

>
> >>>> Yet, I agree that udev tries to set devices too fast.
> >>>
> >>> It tries to set devices exacty when you tell it to do so. It's not like
> >>> it goes trolling for random devices is sysfs.
> >>>
> >>>> It would be better if
> >>>> it would wait for a few milisseconds, to reduce the risk of race conditions.
> >>>
> >>> Gah, I really prefer using properly engineered solutions instead of
> >>> adding crutches.
> >>
> >> I agree.
> >>
> >>>>> And this could be easily added to the udev's keymap utility that is
> >>>>> fired up when we discover evdevX devices.
> >>>>
> >>>> Yes, it can, if you add the IR protocol selection on that tool. A remote
> >>>> controller keycode table has both the protocol and the keycodes.
> >>>> This basically means to merge 99% of the logic inside ir-keytable into the
> >>>> evdev generic tool.
> >>>
> >>> Or just have an utility producing keymap name and feed it as input to
> >>> the generic tools. The way most of utilities work...
> >>
> >> I don't like the idea of running a some logic at udev that would generate
> >> such keymap in runtime just before calling the generic tool. The other
> >
> > Why? You'd just call something like:
> >
> > keymap $name `rc-keymap-name -d $name`
> >
> > where 'keymap' is udev's utility and 'rc-keymap-name' is new utility
> > that incorporates map selection logic currently found in rc-keytable.
> >
> > It looks like format of the keymaps is compatible between 'keymap' and
> > 'ir-keytable' and metadata that is present in your keymaps will not
> > confuse 'keymap' utility.
>
> The format is, currently compatible. However, we'll likely need to change it
> (or to allow the tool to handle also a different format), due to some reasons:
> 1) Protocol and the device name where it is found by default is
> currently a comment;

So? Keep it there, it does not hurt anything.

> 2) We'll need to add a field there specifying the number of the bits
> to be used by the keymap table, in order to use the proper length
> with _V2 ioctls;

You can't calculate this automatically given the length of the
"scancode"? I mean if you have keymap file with

0x12345678936 KEY_COFFEE

that means that scancode is at least 36 bit.

> 3) There are hundreds of keymaps already created for lircd. It
> would be nice to support lircd format, in order to make life
> easier for those that use lirc.

or write a script to convert and be done with it?

>
> If you want to add these to the generic tool, that's fine for me, but, IMO, this
> doesn't sound a good idea. There are already specialized tools for other kinds
> of input devices (mouse: gpm; joystick: jstest; etc). It seems a bad idea
> to merge them into the generic tools.

gpm is different and more akin to X itself and is not limited to mice;
jstest uses differnt interface (jsX, not eventX) so you are right, we
should not mix them into utilities working with event interface.

keymap/ir-keytable both work with event interface.

>
> I think it is better to have a tool that just handle one kind of device, but
> does it well, than trying to extend a generic tool to cover all possible
> devices, and adding lots of caveats there for it to handle all the specifics.
>
> >> alternative (e. g.) to maintain the RC-protocol dependent keytables separate
> >> from the RC protocol used by each table will be a maintenance nightmare.
> >
> > I do not propose splitting keytables, I propose splittign utilities.
> > ir-keytable is a kitchen sink now. It implements 'keymap', 'evtest' and
> > bucnch of other stuff and would be much cleaner if split apart.
>
> It may be broken into a few utilities, by creating a library with the
> common code. I'll think about that when I have some spare time for it.
>

Thanks.

--
Dmitry
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/