Re: [PATCH 1/3] 2.6.14-rc2-mm1: fixes for overflow msec_to_jiffies()

[Willy Tarreau <willy@w.ods.org>]

On Thu, Sep 29, 2005 at 12:52:07PM -0700, Andrew Morton wrote:
> Willy Tarreau <willy@w.ods.org> wrote:
> >
> > Thanks Andrew,
> > 
> > I'm very sorry because I have verified the code with gcc-2.95.3,
> > gcc-3.3.6 and gcc-3.4.4 on x86 and alpha to ensure that everything
> > went smooth on archs where sizeof(long) > sizeof(int). But I've
> > tested all the combinations in user-space for obvious ease of
> > validation. I believe I forgot to use -Wall. What architecture
> > gave you this, and with which compiler please ? I'm willing to
> > fix this as soon as I can understand the root of the problem.
> > 
> 
> http://www.zip.com.au/~akpm/linux/patches/stuff/top-posting.txt

Guess what ? I hate it too when others do it, but I often think that
my mail will be a one-liner which will be easier to read this way, and
of course I'm wrong. Nice FAQ BTW.

> > On Thu, Sep 29, 2005 at 02:43:12AM -0700, Andrew Morton wrote:
> > > Willy Tarreau <willy@w.ods.org> wrote:
> > > >
> > > > +#if HZ <= MSEC_PER_SEC && !(MSEC_PER_SEC % HZ)
> > > >  +#  define MAX_MSEC_OFFSET \
> > > >  +	(ULONG_MAX - (MSEC_PER_SEC / HZ) + 1)
> > > 
> > > That generates numbers which don't fit into unsigned ints, yielding vast
> > > numbers of
> > > 
> > > include/linux/jiffies.h: In function `msecs_to_jiffies':
> > > include/linux/jiffies.h:310: warning: comparison is always false due to limited range of data type
> > > include/linux/jiffies.h: In function `usecs_to_jiffies':
> > > include/linux/jiffies.h:323: warning: comparison is always false due to limited range of data type
> > > 
> 
> This was a ppc64 build, gcc-3.3.3, CONFIG_HZ=250

OK, I have a free account on a ppc64 machine in case I cannot reproduce
on anything else.

> Look a the value which MAX_MSEC_OFFSET will take (it's 2^63 minus a bit). 
> Comparing that to an unsigned int will generate the always-true or
> always-false warning.

I see it. I've done the ulong magic only on the constant computation,
while in theory, I should have casted m to ulong before the comparison
because m is implicitly casted to ulong in the return (common type).
In practise, it should be better to cast MAX_MSEC_OFFSET to unsigned int
in the comparison, but there's still a risk of warning if MAX_MSEC_OFFSET
becomes equal to ~0.

I'll makes a few tests and check that gcc is smart enough to remove the
cast code when unneeded if I cast m to ulong.

Thanks for the details,
Willy