On Thu, Mar 31, 2011 at 11:47:04PM +1000, John Williams wrote:
On Thu, Mar 31, 2011 at 11:23 PM, Wolfram Sang [off-list ref] wrote:
quoted
quoted
Maybe I misunderstand you, in my view it is the responsibility of <vendor>
to create their DTS files to indicate they want <special-card1> to bind to
generic-uio.
Device tree is a OS-neutral hardware description language. "generic-uio"
is neither OS-neutral nor a hardware description. devicetree.org has
more information about this.
If we are trying to be pure, I might argue we are not changing the DTS
language, but rather just add support in Linux for a particular
use-case. There is no violation of DTS syntax.
It might be *recommended* that device trees describe only hardware,
although as Michal points out there is already precedent in the
'chosen' node where this is clearly violated, but in a way that is
compatible with DTS syntax.
There are of course exceptions, particularly for passing boot
information that is OS specific. There is strong pressure to avoid it
however.
Is it forbidden to have DTS descriptions of purely virtual devices, as
would be present if you boot a DTS-driven kernel inside a VM
environment, which provides only virtual implementations of various
devices (ethernet etc)?
'vmware,virt-enet' or whatever?
No, it is not at all forbidden. However it needs to be anchored on a
real implementation of the virtual device. The difference with uio is
that 'uio' is a very specific description of /how/ linux interacts
with the device. It doesn't describe /what/ the interface is.
The virtio stuff is a good example because the interface is defined
indepenently of how Linux actually drives it.
So you could modify the earlier statement to say that device trees
describe only interfaces; not internal OS implementation details.
A really big problem with 'generic-uio' is that it casts a very large
net. If you add 'generic-uio' to a nodes compatible list, then it
immediately precludes any possibility of it being driven by an
in-kernel driver.
However, as already raised there is another way to skin this cat....
quoted
quoted
Our use-case is pretty clear, in FPGA-based systems it is common to create
arbitrary devices that developers just want to control from userspace,
with simple IRQ and IO capabilities (DMA can come later :). �They don't
need to bind to other kernel APIs or subsystems, and don't want to invest
in one-off kernel drivers that simply will never go upstream.
For that, the new_compatible-file would be suitable, I think.
# echo "generic-uio" > /sys/class/uio/<something>
?
Yeah, something like that. I'd prefer something like:
"<vendor>,<hardware-name>" > /sys/bus/platform/drivers/generic-uio/compatible-hardware
That makes it the model to supplement the driver with additional
information about what devices it binds against.
quoted
quoted
UIO is perfect, and simply tagging the device as generic-uio in the DTS is
so simple, clean, and elegant.
Simple, yes (I do understand I wrote the first approach ;)) . Elegant,
not really, because it breaks core conventions of the device tree. For
your case it is a very conveniant hack, but it is still a hack.
Being useful seems like a high priority in the kernel, I'm not ashamed of it! :)
:-)
Regards,
John