Re: [lkp] [gpio] 3c702e9987: kmsg.user_verbs:couldn't_register_device_number

[Michael Welling <mwelling@ieee.org>]

On Sun, Feb 14, 2016 at 11:05:15AM -0800, Greg KH wrote:
> On Sun, Feb 14, 2016 at 06:56:20PM +0100, Linus Walleij wrote:
> > On Sun, Feb 14, 2016 at 6:49 PM, Greg KH <gregkh@linuxfoundation.org> wrote:
> > > On Sun, Feb 14, 2016 at 06:42:11PM +0100, Linus Walleij wrote:
> > >> Greg, heads-up on this... you'd know if this happened
> > >> before.
> > >>
> > >> On Sun, Feb 14, 2016 at 9:06 AM, Michael Welling <mwelling@ieee.org> wrote:
> > >> > On Sun, Feb 14, 2016 at 02:59:06PM +0800, kernel test robot wrote:
> > >> >> FYI, we noticed the below changes on
> > >> >>
> > >> >> https://git.kernel.org/pub/scm/linux/kernel/git/linusw/linux-gpio.git chardev
> > >> >> commit 3c702e9987e261042a07e43460a8148be254412e ("gpio: add a userspace chardev ABI for GPIOs")
> > >> >>
> > >> >>
> > >> >> [    1.951191] user_verbs: couldn't register device number
> > >> >
> > >> > Looks like user_verbs is using a static device node setup.
> > >> >
> > >> > enum {
> > >> >         IB_UVERBS_MAJOR       = 231,
> > >> >         IB_UVERBS_BASE_MINOR  = 192,
> > >> >         IB_UVERBS_MAX_DEVICES = 32
> > >> > };
> > >> >
> > >> > #define IB_UVERBS_BASE_DEV      MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
> > >>
> > >> That's annoying...
> > >> I notice that infiniband is using register_chrdev_region() at
> > >> module_init() time, counting on device major 231 to be free.
> > >
> > > That device major is assigned to Infiniband, why shouldn't it be doing
> > > this?
> > 
> > I mean it's annoying that they collide. (Because of the details I
> > write below, it's fine it's using the assigned number.
> > 
> > > Why not just ask for a new reserved one?  We could give you 261 and
> > > everything should be fine, right?
> > 
> > Sure I can post a patch for that, but it just mitigates the problem.
> > 
> > The report point to the serious problem that on this system
> > some dynamic allocations have already stolen major device
> > numbers 232 thru 255, and 232 and 233 are also assigned.
> > 
> > What do you think about a patch that makes fs/char_dev.c
> > emit a warning when it starts assigning dynamic numbers
> > 233 and below?
> 
> That's fine with me.  I also think maybe we should look into just
> switching all char major/minor allocation to be dynamic, starting at the
> bottom and moving up.  I think the only tools that might have an issue
> with that is the raw device controller, but maybe that has been fixed up
> in userspace, I haven't looked at that in many years.
>

Is there any reason for the CHRDEV_MAJOR_HASH_SIZE being 255?
If we increase the size to say 511 will it break userspace?

In the future I see a robot building a kernel with more that 255 devices and
having to deal with this kind of collision again.

The handling of large major assignment baffles me.
The major numbers outside of the size of the table are just wrapping around to
the beginning again. This is inherently going to cause collisions.

static inline int major_to_index(unsigned major)
{
	return major % CHRDEV_MAJOR_HASH_SIZE;
}

> I thought I had an old patch around somewhere that did that, will go
> look for it this week and see what breaks with it enabled...
> 
> thanks,
> 
> greg k-h