[PATCH 2/2 v3] mmc: dw_mmc: Fix the CTO overflow calculation for 32-bit systems

[Jisheng Zhang]

On Mon, 26 Feb 2018 20:27:22 +0000 Alexey Brodkin wrote:

> Hi Andy,
> 
> On Mon, 2018-02-26 at 20:30 +0200, Andy Shevchenko wrote:
> > On Mon, Feb 26, 2018 at 7:14 PM, Evgeniy Didin
> > <Evgeniy.Didin at synopsys.com> wrote:  
> > > On Mon, 2018-02-26 at 18:53 +0200, Andy Shevchenko wrote:  
> > > > On Mon, Feb 26, 2018 at 5:14 PM, Evgeniy Didin
> > > > <Evgeniy.Didin at synopsys.com> wrote:  
> > > > > On Mon, 2018-02-26 at 16:39 +0200, Andy Shevchenko wrote:  
> > > > > > On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
> > > > > > <Evgeniy.Didin at synopsys.com> wrote:  
> > > > > > > In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
> > > > > > > have been made changes which can cause multiply overflow for 32-bit systems.
> > > > > > > The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
> > > > > > > Lets cast this multiply to u64 type which prevents overflow.
> > > > > > > -       cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> > > > > > > +
> > > > > > > +       cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);  
> > > > > > 
> > > > > > IIRC, someone commented on this or similar, i.e.
> > > > > > 
> > > > > > DIV_ROUND_UP_ULL() ?  
> > > > > 
> > > > > Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
> > > > > because overflow happens on multiply and DIV_ROUND_UP_ULL helps
> > > > > with sum overflow.  
> > > > 
> > > > Did you try to compile your code for 32-bit target?  
> > > 
> > > Yes, we have compiled code for 32-bit system.
> > > I am wondering why are you asking that?  
> > 
> > Because I simple didn't believe you.  
> 
> Well world around us is much more complicated than it sometimes looks like :)
> 
> > ERROR: "__udivdi3" [drivers/mmc/host/dw_mmc.ko] undefined!
> > ...scripts/Makefile.modpost:92: recipe for target '__modpost' failed
> > make[2]: *** [__modpost] Error 1  
> 
> That's right __udivdi3() is not defined for some architectures but it is defined for
> some others including ARC, Xtensa etc, see
> https://elixir.bootlin.com/linux/latest/ident/__udivdi3
> 
> What happens __udivdi3() is implemented in libgcc in one form or another form
> be it pure assembly or generic implementation written in C.
> 
> So maybe we need to add export of __udivdi3() for other arches, what do you think?

Per my understanding, Linux kernel prefer to make use of do_div or implementations
in <linux/math64.h> for 64bit divide

Thanks