On Wed, Dec 06, 2023 at 02:07:16PM +0000, David Laight wrote:
> From: Bjorn Andersson
> > Sent: 06 December 2023 00:44
> >
> > commit 'e84e61bdb97c ("soc: qcom: stats: Add DDR sleep stats")' made it
> > in with a mult_frac() which causes link errors on Arm and PowerPC
> > builds:
> >
> > ERROR: modpost: "__aeabi_uldivmod" [drivers/soc/qcom/qcom_stats.ko] undefined!
> >
> > Expand the mult_frac() to avoid this problem.
> >
> > Fixes: e84e61bdb97c ("soc: qcom: stats: Add DDR sleep stats")
> > Reported-by: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
> > Signed-off-by: Bjorn Andersson <quic_bjorande@xxxxxxxxxxx>
> > ---
> > drivers/soc/qcom/qcom_stats.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/soc/qcom/qcom_stats.c b/drivers/soc/qcom/qcom_stats.c
> > index 4763d62a8cb0..5ba61232313e 100644
> > --- a/drivers/soc/qcom/qcom_stats.c
> > +++ b/drivers/soc/qcom/qcom_stats.c
> > @@ -221,7 +221,8 @@ static int qcom_ddr_stats_show(struct seq_file *s, void *unused)
> >
> > for (i = 0; i < ddr.entry_count; i++) {
> > /* Convert the period to ms */
> > - entry[i].dur = mult_frac(MSEC_PER_SEC, entry[i].dur, ARCH_TIMER_FREQ);
> > + entry[i].dur *= MSEC_PER_SEC;
> > + entry[i].dur = div_u64(entry[i].dur, ARCH_TIMER_FREQ);
>
> Is that right?
> At a guess mult_frac(a, b, c) is doing a 32x32 multiply and then a 64x32
> divide to generate a 32bit result.
> So I'd guess entry[i].dur is 32bit? (this code isn't in -rc4 ...).
> Which means you are now discarding the high bits.
>
entry[i].dur is 64 bit, so this should work just fine.
Arnd proposed that as ARCH_TIMER_FREQ is evenly divisible by
MSEC_PER_SEC we just div_u64(dur, ARCH_TIMER_FREQ / MSEC_PER_SEC), and I
picked that patch instead.
> You've also added a very slow 64bit divide.
Without checking the generated code, I'd expect this to be a slow 64-bit
division already. But this is a debug function, so it should be fine to
take that penalty.
> A multiple by reciprocal calculation will be much better.
> Since absolute accuracy almost certainly doesn't matter here convert:
> dur * 1000 / FREQ
> to
> (dur * (u32)(1000ull << 32 / FREQ)) >> 32
> which will be fine provided FREQ >= 1000
>
I'm quite sure you're right regarding the accuracy. I think as this
isn't in a hot path, the more readable div_u64() feels like a reasonable
choice.
Thank you for your input and suggestion though!
Regards,
Bjorn
> David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
