watchdog ioctl inconsistencies

[Rasmus Villemoes <linux@rasmusvillemoes.dk>]

Hi,

uapi/linux/watchdog.h has these

#define WDIOC_SETOPTIONS        _IOR(WATCHDOG_IOCTL_BASE, 4, int)

This is a write from userspace perspective, so should have been _IOW.

#define WDIOC_KEEPALIVE         _IOR(WATCHDOG_IOCTL_BASE, 5, int)

This one doesn't actually take an argument, so should just have been an
_IO - or if anything, an _IOW. One could be misled to think that if the
int argument has 'V' somewhere (perhaps first or last byte, depending on
endianness) that would count as a magic close.

#define WDIOC_SETTIMEOUT        _IOWR(WATCHDOG_IOCTL_BASE, 6, int)
#define WDIOC_SETPRETIMEOUT     _IOWR(WATCHDOG_IOCTL_BASE, 8, int)

The SETTIMEOUT handling does fall through to the GETTIMEOUT case, so
that one is indeed a "write this, but tell me what value actually took
effect". The SETPRETIMEOUT case ends with a break, so that one is really
_IOW.

There's not much to do about these, I think, but perhaps one could add a
comment to the uapi header containing the magic explains-all phrase
"historical reasons".

Does any static checker actually know about these conventions and peek
inside the _IO*() macros when used as an argument to ioctl(), comparing
the type and constness of the third argument to the direction/type
implied by the macro?

Rasmus