Re: Mostly portable strnlen_user()

[David Miller <davem@davemloft.net>]

From: Linus Torvalds <torvalds@linux-foundation.org>
Date: Fri, 25 May 2012 15:35:38 -0700 (PDT)

> So it's not truly generic, but it's in lib/strnlen_user.c and under the 
> CONFIG_GENERIC_STRNLEN_USER config option, because I'm convinced it should 
> be possible to do something that does the right thing for everybody. 

It should be relatively easy, especially since you work with aligned
words.

> What's the appropriate *clean* interface to allow architectures 
> to tweak the little details?

Maybe something like:

#define HAS_ZERO_CONST_DECLARE	...
#define HAS_ZERO_VAR_DECLARE	...
#define HAS_ZERO_ARGS		...

static inline unsigned long has_zero(unsigned long c ...)
{
}

#define FIND_ZERO_ARGS		...

static inline long find_zero(unsigned long c, unsigned long mask ...)
{
}

So at the use sites it's something like:

long do_strwhatever_user()
{
	HAS_ZERO_CONST_DECLARE;

	...

	while (max >= sizeof(unsigned long)) {
		unsigned long mask, c;
		HAS_ZERO_VAR_DECLARE;

		mask = has_zero(c HAS_ZERO_ARGS);
		if (mask)
			return res + find_zero(c, mask FIND_ZERO_ARGS);
	...
	}
	...
}

A little ugly.  The main problem is the magic args macros.

We can kill off the HAS_ZERO_CONST_DECLARE booger by instead
exposing the magic constants used by the particular has_zero()
implementation.  Then we'll have:

#define HAVE_ZERO_CONST1	REPEAT_BYTE(X)
#define HAVE_ZERO_CONST2	REPEAT_BYTE(X)

static inline unsigned long has_zero(unsigned long c,
				     const unsigned long magic1,
				     const unsigned long magic2);

and then it becomes something like:

long do_strwhatever_user()
{
	const unsigned long magic1 = HAVE_ZERO_CONST1;
	const unsigned long magic2 = HAVE_ZERO_CONST2;

	...

	while (max >= sizeof(unsigned long)) {
		unsigned long mask, c;
		HAS_ZERO_VAR_DECLARE;

		mask = has_zero(c, magic1, magic2);
		if (mask)
			return res + find_zero(c, mask, magic1, magic2);

The remaining problem, which I was trying to accomodate with the funky
ARGS macros, is propagating part of the has_zero() calculation into the
find_zero() call.

Hmmm...