Re: [PATCH v2 next 2/4] lib: mul_u64_u64_div_u64() Use BUG_ON() for divide by zero

[Nicolas Pitre]

On Mon, 19 May 2025, David Laight wrote:

> On Mon, 19 May 2025 08:10:50 +0200
> Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxx> wrote:
> 
> > On Sun, May 18, 2025 at 02:38:46PM +0100, David Laight wrote:
> > > Do an explicit BUG_ON(!divisor) instead of hoping the 'undefined
> > > behaviour' the compiler generated for a compile-time 1/0 is in any
> > > way useful.
> > > 
> > > It may be better to define the function to return ~(u64)0 for
> > > divide by zero.
> > > 
> > > Signed-off-by: David Laight <david.laight.linux@xxxxxxxxx>
> > > ---
> > > 
> > > A new change for v2 of the patchset.
> > > Whereas gcc inserts (IIRC) 'ud2' clang is likely to let the code
> > > continue and generate 'random' results for any 'undefined bahaviour'.
> > > 
> > >  lib/math/div64.c | 10 +++-------
> > >  1 file changed, 3 insertions(+), 7 deletions(-)
> > > 
> > > diff --git a/lib/math/div64.c b/lib/math/div64.c
> > > index a5c966a36836..c426fa0660bc 100644
> > > --- a/lib/math/div64.c
> > > +++ b/lib/math/div64.c
> > > @@ -186,6 +186,9 @@ EXPORT_SYMBOL(iter_div_u64_rem);
> > >  #ifndef mul_u64_u64_div_u64
> > >  u64 mul_u64_u64_div_u64(u64 a, u64 b, u64 d)
> > >  {
> > > +	/* Trigger exception if divisor is zero */
> > > +	BUG_ON(!d);
> > > +  
> > 
> > I'm unsure if I should like the BUG_ON better than return 1/0. My gut
> > feeling is that mul_u64_u64_div_u64() should behave in the same way as
> > e.g. div64_u64 (which is just `return dividend / divisor;` for 64bit
> > archs and thus triggers the same exception as `return 1/0;`.
> 
> You need to execute a run-time 1/0 not a compile-time one.
> clang is likely to decide it is 'undefined behaviour' and just not
> generate any code at all - including removing the 'if (!d)' condition.

The code as it is works perfectly with both gcc and clang: it triggers 
the proper trap or
or call the corresponding exception handler. 

> For x86 gcc does (sometimes at least) generate 'if (!d) asm("ud2")'

And that's perfect. Did you find any occurrence when it is not the case?

> but BUG_ON() adds a table entry for the fault site.

You should really be as close as the behavior you get with a runtime x/y 
where y = 0. When that happens there are no BUG_ON().


Nicolas