Re: [PATCH 3/5] perf_counter: rework ioctl()s

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Tue, 2009-05-12 at 01:58 +0200, Arnd Bergmann wrote:
> On Friday 08 May 2009, Peter Zijlstra wrote:
> > -#define PERF_COUNTER_IOC_ENABLE                _IO ('$', 0)
> > -#define PERF_COUNTER_IOC_DISABLE       _IO ('$', 1)
> > +#define PERF_COUNTER_IOC_ENABLE                _IOW('$', 0, u32)
> > +#define PERF_COUNTER_IOC_DISABLE       _IOW('$', 1, u32)
> >  #define PERF_COUNTER_IOC_REFRESH       _IOW('$', 2, u32)
> > -#define PERF_COUNTER_IOC_RESET         _IO ('$', 3)
> > +#define PERF_COUNTER_IOC_RESET         _IOW('$', 3, u32)
> 
> These ioctl definitions look malformed:
> _IOW('$', 0, u32) means that the ioctl will read a u32 in user
> space pointed to by (u32 __user *)arg, while what perf_ioctl
> actually does is cast arg to a u32. Moreover, exported headers
> should use __u32 instead of u32.
> 
> PERF_COUNTER_IOC_REFRESH apparently was broken already, this
> patch also breaks the other definitions.

Hmm, are you saying that the 3rd argument to unlocked_ioctl is actually
(void __user *) instead of unsigned long?