Re: [PATCH] mtd: nand: s3c2410: fix bug in s3c2410_nand_correct_data()

[Boris Brezillon <boris.brezillon@free-electrons.com>]

Hi Zeng,

On Fri, 8 Apr 2016 13:37:22 +0800
Zeng Zhaoxiu <zhaoxiu.zeng@gmail.com> wrote:

> 在 2016年04月08日 10:18, Boris Brezillon 写道:
> > On Fri, 8 Apr 2016 09:51:04 +0800
> > Zeng Zhaoxiu <zhaoxiu.zeng@gmail.com> wrote:
> >
> >>
> >> 在 2016年04月08日 08:18, Boris Brezillon 写道:
> >>> Hi Zeng,
> >>>
> >>> On Fri,  8 Apr 2016 00:48:17 +0800
> >>> zengzhaoxiu@163.com wrote:
> >>>
> >>>> From: Zeng Zhaoxiu <zhaoxiu.zeng@gmail.com>
> >>>>
> >>>> If there is only one bit difference in the ECC, the function should return 1.
> >>>> The result of "diff0 & ~(1<<fls(diff0))" is equal to diff0, so the function
> >>>> actually returns -1.
> >>>>
> >>>> Here, we can use the simple expression "(diff0 & (diff0 - 1)) == 0" to determine
> >>>> whether the diff0 has only one 1-bit.
> >>> Missing Signed-off-by here.
> >>>
> >>>> ---
> >>>>    drivers/mtd/nand/s3c2410.c | 2 +-
> >>>>    1 file changed, 1 insertion(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/drivers/mtd/nand/s3c2410.c b/drivers/mtd/nand/s3c2410.c
> >>>> index 9c9397b..c9698cf 100644
> >>>> --- a/drivers/mtd/nand/s3c2410.c
> >>>> +++ b/drivers/mtd/nand/s3c2410.c
> >>>> @@ -542,7 +542,7 @@ static int s3c2410_nand_correct_data(struct mtd_info *mtd, u_char *dat,
> >>>>    	diff0 |= (diff1 << 8);
> >>>>    	diff0 |= (diff2 << 16);
> >>>>    
> >>>> -	if ((diff0 & ~(1<<fls(diff0))) == 0)
> >>>> +	if ((diff0 & (diff0 - 1)) == 0)
> >>> Or just
> >>>
> >>> 	if (hweight_long((unsigned long)diff0) == 1)
> >>>
> >>> which is doing exactly what the comment says.
> >>>
> >>> BTW, I don't understand why the current code is wrong? To me, it seems
> >>> it's correctly detecting the case where only a single bit is different.
> >>> What are you trying to fix exactly?
> >>>
> >>> Best Regards,
> >>>
> >>> Boris
> >>>
> >> For example, assuming diff0 is 1, then fls(diff0) is equal to 1, then "~(1 << fls(diff0))" is equal to 0xfffffffd,
> >> then the result of "(diff0 & ~(1 << fls(diff0)))" is 1 , not we expected 0.
> >>
> >> __fls(diff0) and "(fls(diff0) - 1)" are all right, but fls(diff0) is wrong.
> >>
> > Indeed, I forgot that fls() was returning (position + 1). Anyway, I
> > still think using hweight clarifies what you really want to test.
> >
> 
> "(n & (n - 1))" is used in is_power_of_2() in incluse/linux/log2.h,
> it's result is equal to "n & ~(1 << __ffs(n))".
> 
> "(diff & (diff - 1))" is simple and fast, although here is not performance critical.
> To improve readability of this code, we should add a new function and use it.
> 
> /*
>   *  Determine whether some value has more than one 1-bits
>   */
> 
> static inline __attribute__((const))
> bool more_than_1_bit_set(unsigned long n)
> {
>      return (n & (n - 1)) != 0;
> }
> 
> OTOH, I found many determinations like "hweightN(n) > 1" distributed in kernel,
> these determinations are slower than "(n & (n - 1)) != 0" on most CPUs.

Yes, probably, but it may be faster on a few CPUs :). Anyway, not sure
you should bother optimizing this now, especially since this test is in
the ECC correction path, and I doubt it makes any difference (detecting
and correcting errors is what takes most of the time here).

> We can use this new function instead.
> 

In the end, I don't care that much which solution you'll choose, since
it's driver specific code. Pick whatever implementation you prefer and
resend the patch with your SoB.

Thanks,

Boris

-- 
Boris Brezillon, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com