Re: [PATCH] mtd: nand: Fix return type of __DIVIDE() when called with 32-bit

From: Boris Brezillon
Date: Mon May 14 2018 - 07:23:44 EST


On Mon, 14 May 2018 12:49:37 +0200
Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:

> The __DIVIDE() macro checks whether it is called with a 32-bit or 64-bit
> dividend, to select the appropriate divide-and-round-up routine.
> As the check uses the ternary operator, the result will always be
> promoted to a type that can hold both results, i.e. unsigned long long.
>
> When using this result in a division on a 32-bit system, this may lead
> to link errors like:
>
> ERROR: "__udivdi3" [drivers/mtd/nand/raw/nand.ko] undefined!
>
> Fix this by casting the result of the 64-bit division to the type of the
> dividend.
>
> Fixes: 8878b126df769831 ("mtd: nand: add ->exec_op() implementation")
> Signed-off-by: Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>
> ---
> This fixes the root cause of the link failure seen with
> m68k/allmodconfig since commit 3057fcef385348fe ("mtd: rawnand: Make
> sure we wait tWB before polling the STATUS reg").
>
> An alternative mitigation was posted as "[PATCH] m68k: Implement
> ndelay() as an inline function to force type checking/casting"
> (https://lkml.org/lkml/2018/5/13/102).
> ---
> include/linux/mtd/rawnand.h | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/include/linux/mtd/rawnand.h b/include/linux/mtd/rawnand.h
> index 5dad59b312440a9c..d06dc428ea0102ae 100644
> --- a/include/linux/mtd/rawnand.h
> +++ b/include/linux/mtd/rawnand.h
> @@ -871,7 +871,7 @@ struct nand_op_instr {
> #define __DIVIDE(dividend, divisor) ({ \
> sizeof(dividend) == sizeof(u32) ? \
> DIV_ROUND_UP(dividend, divisor) : \
> - DIV_ROUND_UP_ULL(dividend, divisor); \
> + (__typeof__(dividend))DIV_ROUND_UP_ULL(dividend, divisor); \

Hm, it's a bit hard to follow when you place the cast here. One could
wonder why a cast to (__typeof__(dividend)) is needed since
DIV_ROUND_UP_ULL() already returns a (__typeof__(dividend)) type.

How about:

/*
* Cast to type of dividend is needed here to guarantee that the
* result won't be an unsigned long long when the dividend is an
* unsigned long, which is what the compiler does when it sees a
* ternary operator with 2 different return types.
*/
(__typeof__(dividend))(sizeof(dividend) == sizeof(u32) ? \
DIV_ROUND_UP(dividend, divisor) : \
DIV_ROUND_UP_ULL(dividend, divisor));

Actually, I'm not even sure we care about the truncation that could
happen on an unsigned long long -> unsigned long cast because the
delays we express here will anyway be hundreds of nanosecs/millisecs,
so nothing close to the billions of nanosecs/millisecs you can express
with an unsigned long.

So, maybe we should just do:

(unsigned long)(sizeof(dividend) == sizeof(u32) ? \
DIV_ROUND_UP(dividend, divisor) : \
DIV_ROUND_UP_ULL(dividend, divisor));

to make things more readable.

> })
> #define PSEC_TO_NSEC(x) __DIVIDE(x, 1000)
> #define PSEC_TO_MSEC(x) __DIVIDE(x, 1000000000)