Re: [PATCH RFC] UIO: Pass information about ioports to userspace

["Hans J. Koch" <hjk@linutronix.de>]

On Mon, Nov 24, 2008 at 08:11:16PM -0800, Greg KH wrote:
> On Mon, Nov 24, 2008 at 11:19:15AM +0100, Hans J. Koch wrote:
> > On Sun, Nov 23, 2008 at 05:40:54PM -0800, Greg KH wrote:
> > > On Sun, Nov 23, 2008 at 01:14:20PM +0100, Hans J. Koch wrote:
> > > > Devices sometimes have memory where all or parts of it can not be mapped to
> > > > userspace. But it might still be possible to access this memory from
> > > > userspace by other means. An example are PCI cards that advertise not only
> > > > mappable memory but also ioport ranges. On x86 architectures, these can be
> > > > accessed with ioperm, iopl, inb, outb, and friends. Mike Frysinger (CCed)
> > > > reported a similar problem on Blackfin arch where it doesn't seem to be easy
> > > > to mmap non-cached memory but it can still be accessed from userspace.
> > > > 
> > > > This patch allows kernel drivers to pass information about such ports to
> > > > userspace. Similar to the existing mem[] array, it adds a port[] array to
> > > > struct uio_info. Each port range is described by start, size, and porttype.
> > > > 
> > > > If a driver fills in at least one such port range, the UIO core will simply
> > > > pass this information to userspace by creating a new directory "portio"
> > > > underneath /sys/class/uio/uioN/. Similar to the "mem" directory, it will
> > > > contain a subdirectory (portX) for each port range given.
> > > 
> > > This is good, but it would really be nice to provide a way for userspace
> > > to access individual ports without having to have access to all ports in
> > > the system.  Lots of times we don't want to give root privileges to some
> > > programs that only need to read and write simple data to a single
> > > device.
> > 
> > Yes, of course, that'd be nice. But it's very much arch dependent. For
> > example, these x86 ioports need special handling on x86, but you can simply
> > mmap them on powerpc. Port-like memory ranges on other archs might require
> > something completely different.
> > Yes, some generic port access layer would really be good, but I'm not sure
> > if the UIO core is the right place to implement it. Do you already have a
> > solution in mind?
> 
> No I don't, sorry, it's just come up a few times recently and I was
> hoping that you would have something :)

No such luck :) But I'll think about it...

> 
> > Maybe we can look at that in a second step. ATM I just want to avoid these
> > situations where userspace needs ugly tricks to find out which ioports
> > belong to a certain card.
> 
> Agreed, I'm not saying that this patch is not ok at all, sorry if you
> misinterpreted it that way.

I didn't ;-)

> I have no objection to this patch, as long
> as it includes the needed documentation.

Agreed. And I found another flaw. I think that the porttype attribute should
not be an int. A string constant would be much better, what do you think?

I think I'll come up with a second version soon, a new patch and another one
with the documentation.

Thanks,
Hans