Re: [PATCH 1/3] ioctl: generic ioctl dispatcher

[Avi Kivity <avi@redhat.com>]

Andi Kleen wrote:
> Avi Kivity <avi@redhat.com> writes:
>
>   
>> +long dispatch_ioctl(struct inode *inode, struct file *filp,
>> +		    unsigned cmd, unsigned long arg,
>> +		    const struct ioctl_handler *handlers,
>> +		    long (*fallback)(const struct ioctl_arg *arg))
>>     
>
> The basic idea is good, but i don't like the proliferation of callbacks,
> which tends to make complicated code and is ugly for simple code
> (which a lot of ioctls are)
>
>   

If the simple calls mostly don't use the argument as a pointer, they are 
better off using a plain switch.  For my own code, I usually leave the 
boilerplate within the switch and the app-specific code in a separate 
function anyway, so there's no big change in style.

The main motivation here was the extensibility (patch 2), which becomes 
much more difficult with a switch.

> How about you make it return an number that can index a switch() instead?
> Then everything could be still kept in the same function.
>
>   

We need to execute code both before and after the handler, so it would 
look pretty ugly:

long my_ioctl_handler(...)
{
    struct ioctl_arg iarg;
    ...
    long ret;

    ret = dispatch_ioctl_begin(&iarg, ...);
    if (ret < 0)
        return ret;
    switch (ret) {
          case _IOC_KEY(MY_IOCTL):
               // your stuff goes here
               break;
          ...
    }
    dispatch_ioctl_end(&iarg, ret);
    return ret;
}

The only clean way to do this without callbacks is with 
constructors/destructors, but we don't have those in C.

-- 
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.