Re: [RFC] Pushing device/driver binding decisions to userspace

[Ben Collins <ben.collins@ubuntu.com>]

On Sun, 2006-11-12 at 22:45 -0800, Nicholas Miell wrote:
> On Sun, 2006-11-12 at 21:22 -0800, Ben Collins wrote:
> > On Sun, 2006-11-12 at 17:47 -0800, Nicholas Miell wrote:
> > > On Sun, 2006-11-12 at 17:24 -0800, Ben Collins wrote:
> > > > On Sun, 2006-11-12 at 16:49 -0800, Nicholas Miell wrote:
> > > > > On Sun, 2006-11-12 at 15:39 -0800, Ben Collins wrote:
> > > > > 
> > > > > What's wrong with making udev or whatever unbind driver A and then bind
> > > > > driver B if the driver bound by the kernel ends up being the wrong
> > > > > choice? (Besides the inelegance of the kernel choosing one and then
> > > > > userspace immediately choosing the other, of course.)
> > > > > 
> > > > > I'd argue that having multiple drivers for the same hardware is a bit
> > > > > strange to begin with, but that's another issue entirely.
> > > > 
> > > > If two drivers are loaded for the same device, there's no way for udev
> > > > to tell the kernel which driver to use for a device, that I know of.
> > > 
> > > /sys/bus/*/drivers/*/{bind,unbind}
> > 
> > "bind" does not tell the driver core to "bind this device with this
> > driver", it tells it to "bind this driver to whatever devices we match
> > that aren't already bound".
> > 
> > That doesn't solve my use case.
> 
> I don't have any hardware with multiple drivers lying around, but I'm
> fairly certain you can write the bus ID of a device into driver A's
> unbind file and then follow that with a write of that bus ID into driver
> B's bind file and get the effect that you want.

Certainly, that's the case. However, it should be possible to do this
without letting the kernel bind a device to one driver, just to go back
and unbind it, and bind to another driver.

> > 
> > > > Also, that just sounds very horrible to do. If you have udev/dbus events
> > > > flying around for "device present", "device gone", "device present",
> > > > then it could make for a very ugly user experience (think of programs to
> > > > handle devices being started because of these events).
> > > 
> > > So don't fire the events until after the final binding.
> > 
> > It's still not a correct solution. If we want a specific driver to be
> > bound to a specific device, userspace shouldn't have to jump through
> > hoops to do it. It should be simple and clean.
> > 
> > The suggestions you are giving require userspace to work around a
> > deficiency in the kernel, by guessing the ordering requirements to
> > satisfy what the user wants. In cases of hotplugging, it is also
> > sometimes impossible to satisfy these requirements using the current
> > scheme.
> 
> Well, the kernel's deficiency is that there's multiple drivers for the
> same hardware, not that userspace doesn't get first say in how hardware
> is bound to drivers.

No, in a lot of cases, it's not a deficiency. Take the entire
drivers/ata/pata_*.ko list. All of them match an IDE driver, however the
pata driver most times does not support all the same PCI id's for the
matching ide driver.

Also the other case I gave where there is an alsa driver and a media
driver for the same chipset, is by design. It can't be helped. There
actually is a case for wanting one driver or the other in the case of
the exact same hardware, depending on the users desire.