Re: [PATCH] cyber2000fb: New framebuffer_alloc API and class_dev changes

[Russell King <rmk@arm.linux.org.uk>]

On Tue, Sep 16, 2003 at 03:40:09PM +0200, Kronos wrote:
> Il Mon, Sep 15, 2003 at 11:58:32PM +0100, Russell King ha scritto: 
> > On Tue, Sep 16, 2003 at 12:17:42AM +0200, Kronos wrote:
> > > Il Mon, Sep 15, 2003 at 10:40:42PM +0100, Russell King ha scritto: 
> > > > There is another reason why the above is fundamentally flawed - who says
> > > > that "release_cfb_info" will still be in module space by the time you
> > > > need to call it?
> > > > 
> > > > Eg, you unload your framebuffer driver module immediately after the device
> > > > has gone away, but someone is keeping the sysfs files associated with
> > > > the fb_info open.
> > > 
> > > I'm quite sure that this can't happen. If someone is keeping a sysfs
> > > file open module use count won't be zero. Right?
> > 
> > Where are you handling the module use count of the framebuffer driver?
> 
> Every   attribute   (ie.  sysfs   file)   has   a  .owner   field   (see
> include/linux/sysfs.h), the module  use count is handled  by sysfs. If a
> sysfs file is open the module can't be unloaded.

The .owner field is set to the module which defined the sysfs attribute.
This wouldn't be the driver, but would be some other part of the framebuffer
layer.

The first instance I see in the patch at the URL below (fbmem.c):

 static CLASS_DEVICE_ATTR(dev, S_IRUGO, show_dev, NULL);

The owner field would be the module containing fbmem.o, which is separate
from the driver - this doesn't protect the driver code from being unloaded,
and its the driver code which is called via fb_info->release() when this
file is eventually closed.

> > I'm also confused why you want to keep the device IO regions around
> > until all sysfs files have been closed.  I hope you're not thinking
> > of touching the hardware after you've returned from the drivers
> > ->remove method?
> 
> I tought  that some  sysfs file  may use those  things. I was  sure that
> class_dev hold a  reference to pci_dev.dev, but that's  not the case. So
> yes, that's a bad idea. I'll move pci stuff in the unregister function.
> 
> > Is there somewhere I can view the core changes?
> 
> They are on fbdev for review, I put a copy here:
> http://web.tiscali.it/kronoz/linux/fbdev-class_dev-fbmem.c.diff

When I tried to get at the fbdev site, it seemed to be down.  Maybe
some of the kernel documentation needs to be updated?

-- 
Russell King (rmk@arm.linux.org.uk)	http://www.arm.linux.org.uk/personal/
Linux kernel maintainer of:
  2.6 ARM Linux   - http://www.arm.linux.org.uk/
  2.6 PCMCIA      - http://pcmcia.arm.linux.org.uk/
  2.6 Serial core


-------------------------------------------------------
This sf.net email is sponsored by:ThinkGeek
Welcome to geek heaven.
http://thinkgeek.com/sf