Re: V4L2 & request_module("char-major-...")

[Hans Verkuil]

On Saturday 02 August 2008 06:33:50 David Ellingsworth wrote:
> Hans Verkuil wrote:
> > Hi all,
> >
> > I'm in the process of converting v4l2-dev.c (2.6.27, in earlier
> > kernels it was called videodev.c) from using register_chrdev() to
> > using register_chrdev_region() and cdev_add().
> >
> > The problem I have is that register_chrdev provides a
> > file_operations struct. The video_open() function in there performs
> > this piece of code when a video device is opened:
> >
> >         vfl = video_device[minor];
> >         if (vfl == NULL) {
> >                 mutex_unlock(&videodev_lock);
> >                 request_module("char-major-%d-%d", VIDEO_MAJOR,
> > minor); mutex_lock(&videodev_lock);
> >                 vfl = video_device[minor];
> >                 if (vfl == NULL) {
> >                         mutex_unlock(&videodev_lock);
> >                         unlock_kernel();
> >                         return -ENODEV;
> >                 }
> >         }
> >
> > It checks if a V4L2 driver registered this particular minor, if not
> > then it tries to load the appropriate module with a char-major-x-x
> > alias. If still no luck, then we bail out.
> >
> > Now, as I understand it this can only happen if someone used mknod
> > to create device nodes and is not using udev. It's not clear to me
> > however how this can ever work: the char-major alias relies on the
> > fact that someone has to link the minor number with an actual V4L
> > driver, but you do not generally know what minor number will be
> > used by a specific driver (depends on load order, etc). Or am I
> > missing something?
> >
> > My questions are:
> >
> > 1) is this code still relevant?
>
> The code is relevant as long as v4l2-dev allocs and associates a
> block of minor number with it's fops.
>
> > 2) if so, how can I replace this code when I switch to cdev since
> > in that case there is no longer a video_open() that can be used for
> > this.
>
> It's interesting that you mention this, I've examined this same
> problem. Essentially, if v4l2-dev only calls cdev_add on minor
> numbers which have been used, there is no need to keep the
> request_module call in video_open. Since v4l2-dev currently calls
> register_chrdev in videodev_init, cdev_add is always called on the
> entire block. If register_chrdev_region is used instead, cdev_add can
> be called during video_register_device_index and the call to
> request_module can be removed.

OK, clear.

> > 3) I also saw after some googling this proposed patch:
> > http://linux.derkeiler.com/Mailing-Lists/Kernel/2003-09/2925.html
> > It adds a MODULE_ALIAS_CHARDEV_MAJOR line for videodev.c. Either
> > this patch was never applied or quickly removed in videodev.c since
> > it's not there. I'm not sure how this relates to the request_module
> > in the current code and whether it should be added after all or
> > not.
>
> I haven't reviewed this patch, but converting v4l2-dev to use
> register_chrdev_region is rather trivial. However, removing
> video_open can't be done without a) patching char_dev or b) reference
> counting the video_device struct. v4l2-dev currently uses a static
> fops which references video_open. The dependency on v4l2-dev's fops
> can be removed if the cdev struct passed to cdev_add is initialized
> with the the fops provided by drivers which use v4l2-dev. This means
> a cdev struct would be needed for every video_device struct created.
> Embedding the cdev struct in the video_device struct and using
> cdev_init would surely solve this problem,

That was my plan.

> but it exposes another 
> problem.
>
> The problem results from how the cdev struct is reference counted.
> Taking a close look at char_dev you'll see that the reference count
> of the kobject in the cdev struct is initialized during cdev_init or
> cdev_alloc, incremented during cdev_open, and decremented during
> cdev_del and __fputs(file_table.c) using cdev_get and cdev_put. Once
> cdev's kobject reference count reaches 0, cdev_default_release or
> cdev_dynamic_release will be called to cleanup or free the struct
> depending on how it was initialized. Unfortunately,
> cdev_default_release nor cdev_dynamic_release provide a way to
> synchronize freeing of the containing struct. Thus, if a cdev struct
> is embedded into the video_device struct, the video_device struct
> must be reference counted in the same way which essentially ensures
> keeping video_open and the addition of video_close. The alternative
> is to add an optional release callback to the cdev struct and call
> that function during
> cdev_default_release and cdev_dynamic_release so that the containing
> video_device struct may be freed properly.
>
> In addition to the above problem, there is the issue of sysfs. The
> video_device struct has a release callback, but it's use is far from
> correct. Currently this callback _always_ occurs in
> video_unregister_device when unregister_device is called. If a cdev
> struct is embedded in the video_device struct, this would certainly
> result in the freeing of the cdev struct at an inappropriate time.
> Some thought will have to be given about how to handle this properly.
> Personally, I believe the kobject callback (video_release) should
> only free sysfs related structures and not the entire video_device
> struct, like soo many drivers currently do.
>
> I have some patches which address all of the above issues, but I have
> not had time to thoroughly test them and have therefore not yet
> submitted them. I'd be willing to send them to you if you're
> interested.

That would be very welcome. In the meantime I'll dig around some more to 
understand how it works and how it is supposed to work.

Thank you for your input, it's very welcome!

       Hans
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/