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

[Boris Brezillon <boris.brezillon@bootlin.com>]

On Mon, 14 May 2018 14:00:19 +0200
Geert Uytterhoeven <geert@linux-m68k.org> wrote:

> Hi Boris,
> 
> On Mon, May 14, 2018 at 1:46 PM, Boris Brezillon
> <boris.brezillon@bootlin.com> wrote:
> > On Mon, 14 May 2018 13:32:30 +0200
> > Geert Uytterhoeven <geert@linux-m68k.org> wrote:  
> >> On Mon, May 14, 2018 at 1:23 PM, Boris Brezillon
> >> <boris.brezillon@bootlin.com> wrote:  
> >> > On Mon, 14 May 2018 12:49:37 +0200
> >> > Geert Uytterhoeven <geert@linux-m68k.org> 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@linux-m68k.org>
> >> >> ---
> >> >> 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.  
> >>
> >> DIV_ROUND_UP_ULL() does not return __typeof__(dividend), but
> >> unsigned long long.  
> >
> > Except if you entered this branch, that means you passed an unsigned
> > long long dividend (AKA u64), otherwise you would go in DIV_ROUND_UP().
> > Am I missing something?  
> 
> Sure, inside that branch, it does.
> But the compiler considers the whole ternary operator construction, i.e.
> both branches.

Yes, and that's my point. The (__typeof__(dividend)) when placed like
you did is ambiguous. It looks like you're doing a useless cast, while
what you're actually fixing is the case where dividend is an u32 (AKA
unsigned long), and from the reader PoV, the code you're fixing
shouldn't even be reached. Hence my suggestion to move the case one
level up and add a comment ;-).

> 
> >> > 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  
> >>
> >> s/an unsigned long/32-bit/
> >>  
> >> >          * ternary operator with 2 different return types.
> >> >          */
> >> >         (__typeof__(dividend))(sizeof(dividend) == sizeof(u32) ?        \  
> >
> > To be completely safe and handle cases where dividend is an unsigned
> > short or an unsigned, we should probably have:
> >
> >         (__typeof__(dividend))(sizeof(dividend) == sizeof(unsigned long long) ? \  
> 
> "> sizeof(u32)"?  
> 
> /me starts to think about uint128_t...

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

Problem solved :-).

> 
> 
> >                                DIV_ROUND_UP_ULL(dividend, divisor) :
> >                                DIV_ROUND_UP(dividend, divisor));
> >  
> >> >                                DIV_ROUND_UP(dividend, divisor) :        \
> >> >                                DIV_ROUND_UP_ULL(dividend, divisor));  
> >>
> >> Looks fine to me, too.
> >>  
> >> > 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.  
> >>
> >> That would break callers who pass a 64-bit dividend, and expect to receive
> >> a 64-bit quotient back (on 32-bit systems).
> >> Calling e.g. PSEC_TO_NSEC(1000000000000ULL) is valid, passing the
> >> result to ndelay() isn't ;-)  
> >
> > Well, theoretically, yes it's possible, in practice, we only ever pass
> > u32 types to PSEC_TO_NSEC() and u64 types to PSEC_TO_MSEC(), so why
> > bother.  
> 
> True. But __DIVIDE() sounds too generic to add such unobvious limitations.
> 

Is the following version okay?

--->8---
>From 59160e26383e1aca1eafc7078d858b91dd0074ce Mon Sep 17 00:00:00 2001
From: Geert Uytterhoeven <geert@linux-m68k.org>
Date: Mon, 14 May 2018 12:49:37 +0200
Subject: [PATCH] mtd: nand: Fix return type of __DIVIDE() when called with
 32-bit

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 division to the type of the
dividend.

Fixes: 8878b126df769831 ("mtd: nand: add ->exec_op() implementation")
Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Boris Brezillon <boris.brezillon@bootlin.com>
---
 include/linux/mtd/rawnand.h | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/include/linux/mtd/rawnand.h b/include/linux/mtd/rawnand.h
index 5dad59b31244..17c919436f48 100644
--- a/include/linux/mtd/rawnand.h
+++ b/include/linux/mtd/rawnand.h
@@ -867,12 +867,18 @@ struct nand_op_instr {
  * tBERS (during an erase) which all of them are u64 values that cannot be
  * divided by usual kernel macros and must be handled with the special
  * DIV_ROUND_UP_ULL() macro.
+ *
+ * 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 (or smaller),
+ * which is what the compiler does when it sees ternary operator with 2
+ * different return types (picks the largest type to make sure there's no
+ * loss).
  */
-#define __DIVIDE(dividend, divisor) ({					\
-	sizeof(dividend) == sizeof(u32) ?				\
-		DIV_ROUND_UP(dividend, divisor) :			\
-		DIV_ROUND_UP_ULL(dividend, divisor);			\
-		})
+#define __DIVIDE(dividend, divisor) ({						\
+	(__typeof__(dividend))(sizeof(dividend) <= sizeof(unsigned long) ?	\
+			       DIV_ROUND_UP(dividend, divisor) :		\
+			       DIV_ROUND_UP_ULL(dividend, divisor)); 		\
+	})
 #define PSEC_TO_NSEC(x) __DIVIDE(x, 1000)
 #define PSEC_TO_MSEC(x) __DIVIDE(x, 1000000000)