On Fri, 2012-09-28 at 11:41 -0500, Larry Finger wrote:

On 09/28/2012 04:04 AM, David Laight wrote:

quoted

quoted

quoted

Index: wireless-testing-new/drivers/net/wireless/rtlwifi/rtl8192ce/hw.c

[]

quoted

quoted

quoted

-       *(u32 *)&rate_mask = (ratr_bitmap & 0x0fffffff) |
-                                    (ratr_index << 28);
+       for (i = 0; i < 3; i++)
+               rate_mask[i] = ratr_bitmap & (0xff << (i * 4));

[]

quoted

So this either fixes, or adds, an endianness bug.

Yes, the rate_mask array is little endian after this fragment is run, but the 
only use of the byte array is to write it to the device, and LE is what it needs 
no matter the platform. This change fixes an endianness bug.

As I tend to get confused when doing these things, I wrote a small test program 
and ran it on x86_64 and PPC-32 to confirm the result.

Thanks for teaching me about a = b >>= 8. I was not aware that C could do that.

The other thing that could be done is to use cpu_to_le32()
but David's method I think is superior for this use as there's
no absolute guarantee that rate_mask is aligned on a u32.

I'd add parens to make the precedence explicit.

	rate_mask[0] = ratr_bitmap;
	rate_mask[1] = (ratr_bitmap >>= 8);
	rate_mask[2] = (ratr_bitmap >>= 8);
	rate_mask[3] = ((ratr_bitmap >> 8) & 0xf) | (ratr_index << 4);