On Fri, Mar 21, 2003 at 10:08:43AM +0100, Roman Zippel wrote:
> Hi,
>
> On Thu, 20 Mar 2003, Greg KH wrote:
>
> > On Fri, Mar 21, 2003 at 12:03:57AM +0100, Roman Zippel wrote:
> > > I'm unsure how your code will scale. It depends on how that code will be
> > > used. If drivers register a lot of devices, your lookup function has to
> > > scan a possibly very long list of minor devices and that function is
> > > difficult to optimize.
> >
> > And then we grab the BKL :(
>
> This is currently required for either implementation and needs to be moved
> to the driver.
Heh, that is definiatly a 2.7 thing, too many drivers rely on this
functionality :(
> > Hint, optimizing the open() path for char devices is not anything we
> > will probably be doing in 2.6, due to the BKL usage there. It's also
> > not anything anyone has seen on any known benchmarks as a point of
> > contention, so I would not really worry about this for now.
>
> The BKL also shouldn't be a reason to make it unnecessary expensive? I
> don't understand your argument.
I was trying to point out that pre-mature optimiziation of this code
should not be done before we get rid of the most expensive portion, the
bkl. That's all.
> > > char devices don't have partitions, so you hardly need regions. The
> > > problem with the tty layer is that the console and the serial devices
> > > should have different majors.
> >
> > There are a number of char drivers that have "regions". The tty layer
> > support them, and the usb core supports them as two examples. I'm sure
> > there are others. Personally, I like the symmetry with the block device
> > function the way Andries did it.
>
> Every single call to usb_register_dev in 2.5.65 uses exactly 1 minor
> number. Block device drivers need regions because they have partitions
> and we need to find out which device a partition belongs to. Where have
> character devices such requirements?
Oh yeah, I forgot I had cleaned up that api to not reserve minors in
chunks, sorry. It used to do that :)
So only tty drivers currently do this. But that might just be because
it's pretty hard to get a range of minors right now, as the api hasn't
been present. Once we expand the range, I bet it will get quite common
(most character drivers only want from 1-16 minors normally.)
> > > only needed to generate /proc/misc. As soon as character devices are
> > > better integrated into the driver model, even this list is not needed
> > > anymore. This means for simple character devices, we can easily add a
> > > alloc_chardev/add_chardev interface similiar to block devices.
> >
> > No, I don't see /proc/misc going away due to the driver model, I imagine
> > there are too many users of it to disappear. Also, the driver model
> > doesn't care a thing about major/minor numbers so I don't understand how
> > you think it can help in this situation.
>
> I didn't mean that /proc/misc goes away, I meant the misc_list in misc.c.
> They could be other ways to generate /proc/misc.
> /proc/devices, /proc/misc, /proc/tty/drivers, ... is currently mostly
> needed to generate device nodes for dynamic device numbers. This badly
> needs a more generic mechanism.
I agree. But again, 2.7. Remember our feature freeze?
thanks,
greg k-h
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sun Mar 23 2003 - 22:00:40 EST