On Fri, Oct 14, 2011 at 04:57:30PM -0700, Andrew Morton wrote:

On Thu, 06 Oct 2011 15:54:46 +0200
Marek Szyprowski [off-list ref] wrote:

quoted

+#ifdef phys_to_pfn
+/* nothing to do */
+#elif defined __phys_to_pfn
+#  define phys_to_pfn __phys_to_pfn
+#elif defined __va
+#  define phys_to_pfn(x) page_to_pfn(virt_to_page(__va(x)))
+#else
+#  error phys_to_pfn implementation needed
+#endif

Yikes!

This hackery should not be here, please.  If we need a phys_to_pfn()
then let's write a proper one which lives in core MM and arch, then get
it suitably reviewed and integrated and then maintained.

Another question is whether we have any arch where PFN != PHYS >> PAGE_SHIFT?
We've used __phys_to_pfn() to implement that on ARM (with a corresponding
__pfn_to_phys()).  Catalin recently added a cast to __phys_to_pfn() for
LPAE, which I don't think is required:

-#define        __phys_to_pfn(paddr)    ((paddr) >> PAGE_SHIFT)
+#define        __phys_to_pfn(paddr)    ((unsigned long)((paddr) >> PAGE_SHIFT))

since a phys_addr_t >> PAGE_SHIFT will be silently truncated if the passed
in physical address was 64-bit anyway.  (Note: we don't support > 32-bit
PFNs).

So, I'd suggest CMA should just use PFN_DOWN() and be done with it.