The arm defconfig builds failed on today's Linux next tag next-20240304.
Build log:
---------
ERROR: modpost: "__aeabi_uldivmod"
[drivers/gpu/drm/sun4i/sun4i-drm-hdmi.ko] undefined!
Steps to reproduce:
# tuxmake --runtime podman --target-arch arm --toolchain clang-17
--kconfig defconfig LLVM=1 LLVM_IAS=1
Links:
- https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20240304/testrun/22919744/suite/build/test/clang-17-defconfig/history/
- https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20240304/testrun/22919744/suite/build/test/clang-17-defconfig/log
- https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20240304/testrun/22919744/suite/build/test/clang-17-defconfig/details/
--
Linaro LKFT
https://lkft.linaro.org
On Mon, Mar 4, 2024, at 09:07, Naresh Kamboju wrote:
> The arm defconfig builds failed on today's Linux next tag next-20240304.
>
> Build log:
> ---------
> ERROR: modpost: "__aeabi_uldivmod"
> [drivers/gpu/drm/sun4i/sun4i-drm-hdmi.ko] undefined!
>
Apparently caused by the 64-bit division in 358e76fd613a
("drm/sun4i: hdmi: Consolidate atomic_check and mode_valid"):
+static enum drm_mode_status
+sun4i_hdmi_connector_clock_valid(const struct drm_connector *connector,
+ const struct drm_display_mode *mode,
+ unsigned long long clock)
{
- struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
- unsigned long rate = mode->clock * 1000;
- unsigned long diff = rate / 200; /* +-0.5% allowed by HDMI spec */
+ const struct sun4i_hdmi *hdmi = drm_connector_to_sun4i_hdmi(connector);
+ unsigned long diff = clock / 200; /* +-0.5% allowed by HDMI spec */
long rounded_rate;
This used to be a 32-bit division. If the rate is never more than
4.2GHz, clock could be turned back into 'unsigned long' to avoid
the expensive div_u64().
Arnd
On Mon, 04 Mar 2024 12:11:36 +0100
"Arnd Bergmann" <[email protected]> wrote:
Hi,
> On Mon, Mar 4, 2024, at 09:07, Naresh Kamboju wrote:
> > The arm defconfig builds failed on today's Linux next tag next-20240304.
> >
> > Build log:
> > ---------
> > ERROR: modpost: "__aeabi_uldivmod"
> > [drivers/gpu/drm/sun4i/sun4i-drm-hdmi.ko] undefined!
> >
>
> Apparently caused by the 64-bit division in 358e76fd613a
> ("drm/sun4i: hdmi: Consolidate atomic_check and mode_valid"):
>
>
> +static enum drm_mode_status
> +sun4i_hdmi_connector_clock_valid(const struct drm_connector *connector,
> + const struct drm_display_mode *mode,
> + unsigned long long clock)
> {
> - struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
> - unsigned long rate = mode->clock * 1000;
> - unsigned long diff = rate / 200; /* +-0.5% allowed by HDMI spec */
> + const struct sun4i_hdmi *hdmi = drm_connector_to_sun4i_hdmi(connector);
> + unsigned long diff = clock / 200; /* +-0.5% allowed by HDMI spec */
Wouldn't "div_u64(clock, 200)" solve this problem?
Cheers,
Andre
> long rounded_rate;
>
> This used to be a 32-bit division. If the rate is never more than
> 4.2GHz, clock could be turned back into 'unsigned long' to avoid
> the expensive div_u64().
>
> Arnd
>
On Mon, Mar 4, 2024, at 12:24, Andre Przywara wrote:
> On Mon, 04 Mar 2024 12:11:36 +0100 "Arnd Bergmann" <[email protected]> wrote:
>>
>> This used to be a 32-bit division. If the rate is never more than
>> 4.2GHz, clock could be turned back into 'unsigned long' to avoid
>> the expensive div_u64().
>
> Wouldn't "div_u64(clock, 200)" solve this problem?
Yes, that's why I mentioned it as the worse of the two obvious
solutions. ;-)
Arnd
On Mon, 04 Mar 2024 12:26:46 +0100
"Arnd Bergmann" <[email protected]> wrote:
> On Mon, Mar 4, 2024, at 12:24, Andre Przywara wrote:
> > On Mon, 04 Mar 2024 12:11:36 +0100 "Arnd Bergmann" <[email protected]> wrote:
> >>
> >> This used to be a 32-bit division. If the rate is never more than
> >> 4.2GHz, clock could be turned back into 'unsigned long' to avoid
> >> the expensive div_u64().
> >
> > Wouldn't "div_u64(clock, 200)" solve this problem?
>
> Yes, that's why I mentioned it as the worse of the two obvious
> solutions. ;-)
Argh, should have cleaned my glasses first ;-)
I guess I was put somehow put off by the word "expensive". While it's
admittedly not trivial, I wonder if we care about the (hidden) complexity
of that function? I mean it's neither core code nor something called
frequently?
I don't think we have any clock exceeding 3GHz at the moment, but it
sounds fishy to rely on that.
Cheers,
Andre
Hi,
On Mon, Mar 04, 2024 at 12:11:36PM +0100, Arnd Bergmann wrote:
> On Mon, Mar 4, 2024, at 09:07, Naresh Kamboju wrote:
> > The arm defconfig builds failed on today's Linux next tag next-20240304.
> >
> > Build log:
> > ---------
> > ERROR: modpost: "__aeabi_uldivmod"
> > [drivers/gpu/drm/sun4i/sun4i-drm-hdmi.ko] undefined!
> >
>
> Apparently caused by the 64-bit division in 358e76fd613a
> ("drm/sun4i: hdmi: Consolidate atomic_check and mode_valid"):
>
>
> +static enum drm_mode_status
> +sun4i_hdmi_connector_clock_valid(const struct drm_connector *connector,
> + const struct drm_display_mode *mode,
> + unsigned long long clock)
> {
> - struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
> - unsigned long rate = mode->clock * 1000;
> - unsigned long diff = rate / 200; /* +-0.5% allowed by HDMI spec */
> + const struct sun4i_hdmi *hdmi = drm_connector_to_sun4i_hdmi(connector);
> + unsigned long diff = clock / 200; /* +-0.5% allowed by HDMI spec */
> long rounded_rate;
>
> This used to be a 32-bit division. If the rate is never more than
> 4.2GHz, clock could be turned back into 'unsigned long' to avoid
> the expensive div_u64().
I sent a fix for it this morning:
https://lore.kernel.org/r/[email protected]
The framework will pass an unsigned long long because HDMI character
rates can go up to 5.9GHz.
Maxime
On Mon, Mar 4, 2024, at 12:45, Andre Przywara wrote:
> On Mon, 04 Mar 2024 12:26:46 +0100
> "Arnd Bergmann" <[email protected]> wrote:
>
>> On Mon, Mar 4, 2024, at 12:24, Andre Przywara wrote:
>> > On Mon, 04 Mar 2024 12:11:36 +0100 "Arnd Bergmann" <[email protected]> wrote:
>> >>
>> >> This used to be a 32-bit division. If the rate is never more than
>> >> 4.2GHz, clock could be turned back into 'unsigned long' to avoid
>> >> the expensive div_u64().
>> >
>> > Wouldn't "div_u64(clock, 200)" solve this problem?
>>
>> Yes, that's why I mentioned it as the worse of the two obvious
>> solutions. ;-)
>
> Argh, should have cleaned my glasses first ;-)
>
> I guess I was put somehow put off by the word "expensive". While it's
> admittedly not trivial, I wonder if we care about the (hidden) complexity
> of that function? I mean it's neither core code nor something called
> frequently?
It's not critical if this is called infrequently, and as Maxime
just replied, the 64-bit division is in fact required here.
Since we are dividing by a constant value (200), there is a good
chance that this will be get turned into fairly efficient
multiply/shift code.
> I don't think we have any clock exceeding 3GHz at the moment, but it
> sounds fishy to rely on that.
Right, it's just important to look at each case individually.
The cost of 64-bit division is crazy if it gets called repeatedly,
which is of course the entire reason we don't provide a
__aeabi_uldivmod() function and require developers to think
before adding div_u64().
Arnd
On Mon, 4 Mar 2024 at 13:35, Arnd Bergmann <[email protected]> wrote:
>
> On Mon, Mar 4, 2024, at 12:45, Andre Przywara wrote:
> > On Mon, 04 Mar 2024 12:26:46 +0100
> > "Arnd Bergmann" <[email protected]> wrote:
> >
> >> On Mon, Mar 4, 2024, at 12:24, Andre Przywara wrote:
> >> > On Mon, 04 Mar 2024 12:11:36 +0100 "Arnd Bergmann" <[email protected]> wrote:
> >> >>
> >> >> This used to be a 32-bit division. If the rate is never more than
> >> >> 4.2GHz, clock could be turned back into 'unsigned long' to avoid
> >> >> the expensive div_u64().
> >> >
> >> > Wouldn't "div_u64(clock, 200)" solve this problem?
> >>
> >> Yes, that's why I mentioned it as the worse of the two obvious
> >> solutions. ;-)
> >
> > Argh, should have cleaned my glasses first ;-)
> >
> > I guess I was put somehow put off by the word "expensive". While it's
> > admittedly not trivial, I wonder if we care about the (hidden) complexity
> > of that function? I mean it's neither core code nor something called
> > frequently?
>
> It's not critical if this is called infrequently, and as Maxime
> just replied, the 64-bit division is in fact required here.
> Since we are dividing by a constant value (200), there is a good
> chance that this will be get turned into fairly efficient
> multiply/shift code.
>
Clang does not implement that optimization for 64-bit division. That
is how we ended up with this error in the first place.
Perhaps it is worthwhile to make div_u64() check its divisor, e.g.,
--- a/include/linux/math64.h
+++ b/include/linux/math64.h
@@ -127,6 +127,9 @@
static inline u64 div_u64(u64 dividend, u32 divisor)
{
u32 remainder;
+
+ if (IS_ENABLED(CONFIG_CC_IS_GCC) && __builtin_constant_p(divisor))
+ return dividend / divisor;
return div_u64_rem(dividend, divisor, &remainder);
}
#endif
On Mon, Mar 4, 2024, at 14:01, Ard Biesheuvel wrote:
> On Mon, 4 Mar 2024 at 13:35, Arnd Bergmann <[email protected]> wrote:
>> On Mon, Mar 4, 2024, at 12:45, Andre Przywara wrote:
>> It's not critical if this is called infrequently, and as Maxime
>> just replied, the 64-bit division is in fact required here.
>> Since we are dividing by a constant value (200), there is a good
>> chance that this will be get turned into fairly efficient
>> multiply/shift code.
>>
>
> Clang does not implement that optimization for 64-bit division. That
> is how we ended up with this error in the first place.
I meant it will use the optimization after the patch to convert
the plain '/' to div_u64().
> Perhaps it is worthwhile to make div_u64() check its divisor, e.g.,
>
> --- a/include/linux/math64.h
> +++ b/include/linux/math64.h
> @@ -127,6 +127,9 @@
> static inline u64 div_u64(u64 dividend, u32 divisor)
> {
> u32 remainder;
> +
> + if (IS_ENABLED(CONFIG_CC_IS_GCC) && __builtin_constant_p(divisor))
> + return dividend / divisor;
> return div_u64_rem(dividend, divisor, &remainder);
> }
I think the div_u64()->do_div()->__div64_const32()->__arch_xprod_64()
optimization in asm-generic/div64.h already produces what we want
on both compilers. Is there something missing there?
Arnd
On Mon, 4 Mar 2024 at 14:49, Arnd Bergmann <[email protected]> wrote:
>
> On Mon, Mar 4, 2024, at 14:01, Ard Biesheuvel wrote:
> > On Mon, 4 Mar 2024 at 13:35, Arnd Bergmann <[email protected]> wrote:
> >> On Mon, Mar 4, 2024, at 12:45, Andre Przywara wrote:
> >> It's not critical if this is called infrequently, and as Maxime
> >> just replied, the 64-bit division is in fact required here.
> >> Since we are dividing by a constant value (200), there is a good
> >> chance that this will be get turned into fairly efficient
> >> multiply/shift code.
> >>
> >
> > Clang does not implement that optimization for 64-bit division. That
> > is how we ended up with this error in the first place.
>
> I meant it will use the optimization after the patch to convert
> the plain '/' to div_u64().
>
Ah ok.
I did not realize we implement the same optimization in our code as
the one that GCC will apply when encountering a compile-time constant
divisor.
> > Perhaps it is worthwhile to make div_u64() check its divisor, e.g.,
> >
> > --- a/include/linux/math64.h
> > +++ b/include/linux/math64.h
> > @@ -127,6 +127,9 @@
> > static inline u64 div_u64(u64 dividend, u32 divisor)
> > {
> > u32 remainder;
> > +
> > + if (IS_ENABLED(CONFIG_CC_IS_GCC) && __builtin_constant_p(divisor))
> > + return dividend / divisor;
> > return div_u64_rem(dividend, divisor, &remainder);
> > }
>
> I think the div_u64()->do_div()->__div64_const32()->__arch_xprod_64()
> optimization in asm-generic/div64.h already produces what we want
> on both compilers. Is there something missing there?
>
No, you are right. I thought we were relying on GCC for the optimization here.
From: Maxime Ripard
> Sent: 04 March 2024 11:46
>
> On Mon, Mar 04, 2024 at 12:11:36PM +0100, Arnd Bergmann wrote:
> > On Mon, Mar 4, 2024, at 09:07, Naresh Kamboju wrote:
> > > The arm defconfig builds failed on today's Linux next tag next-20240304.
> > >
> > > Build log:
> > > ---------
> > > ERROR: modpost: "__aeabi_uldivmod"
> > > [drivers/gpu/drm/sun4i/sun4i-drm-hdmi.ko] undefined!
> > >
> >
> > Apparently caused by the 64-bit division in 358e76fd613a
> > ("drm/sun4i: hdmi: Consolidate atomic_check and mode_valid"):
> >
> >
> > +static enum drm_mode_status
> > +sun4i_hdmi_connector_clock_valid(const struct drm_connector *connector,
> > + const struct drm_display_mode *mode,
> > + unsigned long long clock)
> > {
> > - struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
> > - unsigned long rate = mode->clock * 1000;
> > - unsigned long diff = rate / 200; /* +-0.5% allowed by HDMI spec */
> > + const struct sun4i_hdmi *hdmi = drm_connector_to_sun4i_hdmi(connector);
> > + unsigned long diff = clock / 200; /* +-0.5% allowed by HDMI spec */
> > long rounded_rate;
> >
> > This used to be a 32-bit division. If the rate is never more than
> > 4.2GHz, clock could be turned back into 'unsigned long' to avoid
> > the expensive div_u64().
>
> I sent a fix for it this morning:
> https://lore.kernel.org/r/[email protected]
>
> The framework will pass an unsigned long long because HDMI character
> rates can go up to 5.9GHz.
You could do:
/* The max clock is 5.9GHz, split the divide */
u32 diff = (u32)(clock / 8) / (200/8);
The code should really use u32 and u64.
Otherwise the sizes are different on 32bit.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
On Sat, Mar 9, 2024 at 3:34 PM David Laight <[email protected]> wrote:
> From: Maxime Ripard
> > Sent: 04 March 2024 11:46
> >
> > On Mon, Mar 04, 2024 at 12:11:36PM +0100, Arnd Bergmann wrote:
> > > On Mon, Mar 4, 2024, at 09:07, Naresh Kamboju wrote:
> > > > The arm defconfig builds failed on today's Linux next tag next-20240304.
> > > >
> > > > Build log:
> > > > ---------
> > > > ERROR: modpost: "__aeabi_uldivmod"
> > > > [drivers/gpu/drm/sun4i/sun4i-drm-hdmi.ko] undefined!
> > > >
> > >
> > > Apparently caused by the 64-bit division in 358e76fd613a
> > > ("drm/sun4i: hdmi: Consolidate atomic_check and mode_valid"):
> > >
> > >
> > > +static enum drm_mode_status
> > > +sun4i_hdmi_connector_clock_valid(const struct drm_connector *connector,
> > > + const struct drm_display_mode *mode,
> > > + unsigned long long clock)
> > > {
> > > - struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
> > > - unsigned long rate = mode->clock * 1000;
> > > - unsigned long diff = rate / 200; /* +-0.5% allowed by HDMI spec */
> > > + const struct sun4i_hdmi *hdmi = drm_connector_to_sun4i_hdmi(connector);
> > > + unsigned long diff = clock / 200; /* +-0.5% allowed by HDMI spec */
> > > long rounded_rate;
> > >
> > > This used to be a 32-bit division. If the rate is never more than
> > > 4.2GHz, clock could be turned back into 'unsigned long' to avoid
> > > the expensive div_u64().
> >
> > I sent a fix for it this morning:
> > https://lore.kernel.org/r/[email protected]
> >
> > The framework will pass an unsigned long long because HDMI character
> > rates can go up to 5.9GHz.
>
> You could do:
> /* The max clock is 5.9GHz, split the divide */
> u32 diff = (u32)(clock / 8) / (200/8);
+1, as the issue is still present in current next, as per the recent
nagging from the build bots.
> The code should really use u32 and u64.
> Otherwise the sizes are different on 32bit.
The code is already using a variety of types (long, unsigned long long,
unsigned long) :-(
max_t(unsigned long, rounded_rate, clock) -
min_t(unsigned long, rounded_rate, clock) < diff)
At least u64 should make it very clear clock does not fit in 32-bit.
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68korg
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
On Thu, Mar 14, 2024 at 10:27 AM Geert Uytterhoeven
<[email protected]> wrote:
> On Sat, Mar 9, 2024 at 3:34 PM David Laight <[email protected]> wrote:
> > From: Maxime Ripard
> > > Sent: 04 March 2024 11:46
> > >
> > > On Mon, Mar 04, 2024 at 12:11:36PM +0100, Arnd Bergmann wrote:
> > > > On Mon, Mar 4, 2024, at 09:07, Naresh Kamboju wrote:
> > > > > The arm defconfig builds failed on today's Linux next tag next-20240304.
> > > > >
> > > > > Build log:
> > > > > ---------
> > > > > ERROR: modpost: "__aeabi_uldivmod"
> > > > > [drivers/gpu/drm/sun4i/sun4i-drm-hdmi.ko] undefined!
> > > > >
> > > >
> > > > Apparently caused by the 64-bit division in 358e76fd613a
> > > > ("drm/sun4i: hdmi: Consolidate atomic_check and mode_valid"):
> > > >
> > > >
> > > > +static enum drm_mode_status
> > > > +sun4i_hdmi_connector_clock_valid(const struct drm_connector *connector,
> > > > + const struct drm_display_mode *mode,
> > > > + unsigned long long clock)
> > > > {
> > > > - struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
> > > > - unsigned long rate = mode->clock * 1000;
> > > > - unsigned long diff = rate / 200; /* +-0.5% allowed by HDMI spec */
> > > > + const struct sun4i_hdmi *hdmi = drm_connector_to_sun4i_hdmi(connector);
> > > > + unsigned long diff = clock / 200; /* +-0.5% allowed by HDMI spec */
> > > > long rounded_rate;
> > > >
> > > > This used to be a 32-bit division. If the rate is never more than
> > > > 4.2GHz, clock could be turned back into 'unsigned long' to avoid
> > > > the expensive div_u64().
> > >
> > > I sent a fix for it this morning:
> > > https://lore.kernel.org/r/[email protected]
> > >
> > > The framework will pass an unsigned long long because HDMI character
> > > rates can go up to 5.9GHz.
> >
> > You could do:
> > /* The max clock is 5.9GHz, split the divide */
> > u32 diff = (u32)(clock / 8) / (200/8);
>
> +1, as the issue is still present in current next, as per the recent
> nagging from the build bots.
Oops, s/next/upstream/.
Well, less 32-bit build testing for the next few days :-(
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68korg
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
On Thu, Mar 14, 2024 at 10:27:23AM +0100, Geert Uytterhoeven wrote:
> On Sat, Mar 9, 2024 at 3:34 PM David Laight <[email protected]> wrote:
> > From: Maxime Ripard
> > > Sent: 04 March 2024 11:46
> > >
> > > On Mon, Mar 04, 2024 at 12:11:36PM +0100, Arnd Bergmann wrote:
> > > > On Mon, Mar 4, 2024, at 09:07, Naresh Kamboju wrote:
> > > > > The arm defconfig builds failed on today's Linux next tag next-20240304.
> > > > >
> > > > > Build log:
> > > > > ---------
> > > > > ERROR: modpost: "__aeabi_uldivmod"
> > > > > [drivers/gpu/drm/sun4i/sun4i-drm-hdmi.ko] undefined!
> > > > >
> > > >
> > > > Apparently caused by the 64-bit division in 358e76fd613a
> > > > ("drm/sun4i: hdmi: Consolidate atomic_check and mode_valid"):
> > > >
> > > >
> > > > +static enum drm_mode_status
> > > > +sun4i_hdmi_connector_clock_valid(const struct drm_connector *connector,
> > > > + const struct drm_display_mode *mode,
> > > > + unsigned long long clock)
> > > > {
> > > > - struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
> > > > - unsigned long rate = mode->clock * 1000;
> > > > - unsigned long diff = rate / 200; /* +-0.5% allowed by HDMI spec */
> > > > + const struct sun4i_hdmi *hdmi = drm_connector_to_sun4i_hdmi(connector);
> > > > + unsigned long diff = clock / 200; /* +-0.5% allowed by HDMI spec */
> > > > long rounded_rate;
> > > >
> > > > This used to be a 32-bit division. If the rate is never more than
> > > > 4.2GHz, clock could be turned back into 'unsigned long' to avoid
> > > > the expensive div_u64().
> > >
> > > I sent a fix for it this morning:
> > > https://lore.kernel.org/r/[email protected]
> > >
> > > The framework will pass an unsigned long long because HDMI character
> > > rates can go up to 5.9GHz.
> >
> > You could do:
> > /* The max clock is 5.9GHz, split the divide */
> > u32 diff = (u32)(clock / 8) / (200/8);
>
> +1, as the issue is still present in current next, as per the recent
> nagging from the build bots.
A patch to fix it has been merged today and will show up tomorrow in next.
Maxime