On 19 Apr 1999, Thomas Bushnell, BSG wrote: > OKUJI Yoshinori <[EMAIL PROTECTED]> writes: > > > The attempt I once began foundered on the lack of adequate kernel > > > facilities to do it, but that's fixable. > > I want to know what facilities are lack precisely. If you give me > > the clue, I may fix that. > Basically, we need a call that will return all the valid device names > that the kernel understands (the list should reflect what is actually > present on the machine, and not be a list just of particular pieces of > hardware that might be there). > > It also needs to be possible to tell, for each device, exactly what > kind of device it is, in some manner that does not cause the device to > do anything at all. (Doing device_open/device_get_status/device_close > would be OK, except that some devices block on open, and many do > things on close, and there is no guarantee that every device will > support an appropriate device_get_status type query.)
Would having such a device structure (kernel level devices exporting symbols) be all that much of an architecture win? I still feel that having the kernel only export hardware access (providing a namespace where a device can attach to a file in that namespace and gain hardware access, all controlled via standard hurd authentication) would be a big win. The ability for drivers to be userland, indeed, for a driver to run on a machine other than the one the hardware device is on, provides significant benefits.
