Re: Broken ioctl error returns (was Re: [PATCH 2/3] block: fail SCSIpassthrough ioctls on partition devices)

[Mauro Carvalho Chehab]

On 05-01-2012 15:47, Linus Torvalds wrote:
> On Thu, Jan 5, 2012 at 9:37 AM, Mauro Carvalho Chehab
> <mchehab@xxxxxxxxxxxxx> wrote:
>>
>> For the media drivers, we've already fixed it, at the V4L side:
>> -EINVAL doesn't mean that an ioctl is not supported anymore.
>> I think that such fix went into Kernel 3.1.
> 
> Ok, I'm happy to hear that the thing should be fixed. My grepping
> still found a fair amount of EINVAL returns both in code and in the
> Documentation subdirectory for media ioctls, but it really was just
> grepping with a few lines of context, so I didn't look closer at the
> semantics. I was just looking for certain patterns (ie grepping for
> "EINVAL" near ioctl or ENOIOCTLCMD etc) that I thought might indicate
> problem spots, and the media subdirectory had a lot of them.

Yeah, there are lots of EINVAL there, as the API is fairly complex
(about 80 ioctl's for V4L, plus a similar set for DVB),  but there's
an ioctl dispatcher inside the V4L core that handles the ENOTTY case,
at drivers/media/video/v4l2-ioctl.c.

You'll see some -EINVAL things there, but they're due to errors
on parameters (the semantics there is somewhat complex, to avoid
returning -ENOTTY where a -EINVAL should be returned, instead).

In summary, the code there is:

static long __video_do_ioctl(struct file *file,
                unsigned int cmd, void *arg)
{
...
        long ret = -ENOTTY;
...
	switch (cmd) {
		/*
		 * several ioctl callbacks here. if they're not
		 * implemented, break (e. g. -ENOTTY will be returned).
		 */
...
	}
...
        return ret;

The context there is too big for noticing it with a few lines of
context, and too complex as well, as some ioctl's may be implemented 
by more than one callback, depending on what's needed, and some 
others have a default implementation there. This is somewhat similar 
to file ops callbacks.

> Can you test the patch with some media capture apps (preferably with
> the obvious fix for the problem that Paulo already pointed out -
> although that won't actually matter until some block driver starts
> using ENOIOCTLCMD there, so even the unfixed patch should mostly work
> for testing)?

Sure. I'm currently traveling, so I have just my "first aids kit" of devices
but they should be enough for testing it. I'll return you as soon as I finish
compiling the kernel on this slow 4 years-old notebook and run some
tests with the usual applications.

Regards,
Mauro
--
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/