[PATCH v2 0/3] PBL: enable timeouts in read_poll_timeout macros

[PATCH v2 0/3] PBL: enable timeouts in read_poll_timeout macros

[Stefan Kerkmann]

Without a time source the read_poll_timeout functions will deadlock in the PBL
if the break condition is never met. This series introduces the necessary
timing functions in the PBL for ARMv7 and ARMv8 based on the ARM architected
timer and enable their usage if available.

Signed-off-by: Stefan Kerkmann <s.kerkmann@pengutronix.de>
---
Changes in v2:
- Strictly guard the pbl timer support on 32bit ARM for ARMv7 only builds
- Link to v1: https://lore.kernel.org/r/20250121-feature-pbl-get-time-ns-v1-0-c3d493397846@pengutronix.de

To: Sascha Hauer <s.hauer@pengutronix.de>
To: "open list:BAREBOX" <barebox@lists.infradead.org>

---
Stefan Kerkmann (3):
      ARM64: lib64: pbl: implement get_time_ns and is_timeout
      ARM: lib32: pbl: implement get_time_ns and is_timeout
      pbl: introduce HAS_PBL_CLOCKSOURCE marker

 arch/arm/cpu/Kconfig                   |  2 ++
 arch/arm/lib32/Makefile                |  2 +-
 arch/arm/lib32/arm_architected_timer.c | 15 ++++++++++++++-
 arch/arm/lib64/delay.c                 | 17 +++++++++++++++--
 include/linux/iopoll.h                 |  2 +-
 pbl/Kconfig                            |  3 +++
 6 files changed, 36 insertions(+), 5 deletions(-)
---
base-commit: 4705656eeeaba0dd3617b69172328daf4dbf9060
change-id: 20250121-feature-pbl-get-time-ns-6d9e8874d582

Best regards,
--  
Stefan Kerkmann <s.kerkmann@pengutronix.de>

[PATCH v2 1/3] ARM64: lib64: pbl: implement get_time_ns and is_timeout

[Stefan Kerkmann]

All ARMv8-A cores implement the 64bit wide generic timer CNTPCT[1], thus
we can use it to implement the get_time_ns and is_timeout helpers which
in turn enable the whole read_poll_timeout class of functions in the
PBL. As it is guaranteed that the timer will not wrap for 40 years no
overflow handling is necessary.

[1]: See "ARM Architecture Reference Manual for A-profile architecture
(rev L.a)", chapter D12 "The Generic Timer in AArch64 state"

Signed-off-by: Stefan Kerkmann <s.kerkmann@pengutronix.de>
---
 arch/arm/lib64/delay.c | 17 +++++++++++++++--
 1 file changed, 15 insertions(+), 2 deletions(-)

diff --git a/arch/arm/lib64/delay.c b/arch/arm/lib64/delay.c
index 78eab33f8d..00b20c5db1 100644
--- a/arch/arm/lib64/delay.c
+++ b/arch/arm/lib64/delay.c
@@ -7,7 +7,7 @@
 void udelay(unsigned long us)
 {
 	unsigned long cntfrq = get_cntfrq();
-	unsigned long ticks = (us * cntfrq) / 1000000;
+	unsigned long ticks = (us * cntfrq) / MSECOND;
 	unsigned long start = get_cntpct();
 
 	while ((long)(start + ticks - get_cntpct()) > 0);
@@ -15,5 +15,18 @@ void udelay(unsigned long us)
 
 void mdelay(unsigned long ms)
 {
-	udelay(ms * 1000);
+	udelay(ms * USECOND);
+}
+
+uint64_t get_time_ns(void)
+{
+	return get_cntpct() * SECOND / get_cntfrq();
+}
+
+int is_timeout(uint64_t start, uint64_t time_offset_ns)
+{
+	if ((int64_t)(start + time_offset_ns - get_time_ns()) < 0)
+		return 1;
+	else
+		return 0;
 }

-- 
2.47.3

[PATCH v2 2/3] ARM: lib32: pbl: implement get_time_ns and is_timeout

[Stefan Kerkmann]

ARMv7-A cores can implement the optional 64bit wide generic timer
CNTPCT[1]. If it is present we can use it to implement the get_time_ns
and is_timeout helpers which in turn enable the whole read_poll_timeout
class of functions in the PBL. As it is guaranteed that the timer won't
wrap for 40 years no overflow handling is necessary.

[1]: See "ARM Architecture Reference Manual ARMv7-A and ARMv7-R edition
(rev C.d)", Chapter B8 "The Generic Timer"

Signed-off-by: Stefan Kerkmann <s.kerkmann@pengutronix.de>
---
 arch/arm/lib32/arm_architected_timer.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/arch/arm/lib32/arm_architected_timer.c b/arch/arm/lib32/arm_architected_timer.c
index 54eca13f8b..75f74cbc15 100644
--- a/arch/arm/lib32/arm_architected_timer.c
+++ b/arch/arm/lib32/arm_architected_timer.c
@@ -10,8 +10,21 @@ void arm_architected_timer_udelay(unsigned long us)
 	unsigned long long ticks, cntfrq = get_cntfrq();
 	unsigned long long start = get_cntpct();
 
-	ticks = DIV_ROUND_DOWN_ULL((us * cntfrq), 1000000);
+	ticks = DIV_ROUND_DOWN_ULL((us * cntfrq), MSECOND);
 
 	while ((long)(start + ticks - get_cntpct()) > 0)
 		;
 }
+
+uint64_t get_time_ns(void)
+{
+	return get_cntpct() * SECOND / get_cntfrq();
+}
+
+int is_timeout(uint64_t start, uint64_t time_offset_ns)
+{
+	if ((int64_t)(start + time_offset_ns - get_time_ns()) < 0)
+		return 1;
+	else
+		return 0;
+}

-- 
2.47.3

[PATCH v2 3/3] pbl: introduce HAS_PBL_CLOCKSOURCE marker

[Stefan Kerkmann]

This finally enables the usage of the polled timeout functions in the
barebox pbl with real timeouts for supported architectures. Currently
only ARMv7 and AARCH64 are enabled.

Signed-off-by: Stefan Kerkmann <s.kerkmann@pengutronix.de>
---
 arch/arm/cpu/Kconfig    | 2 ++
 arch/arm/lib32/Makefile | 2 +-
 include/linux/iopoll.h  | 2 +-
 pbl/Kconfig             | 3 +++
 4 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/arch/arm/cpu/Kconfig b/arch/arm/cpu/Kconfig
index 47c116ea98..684c7eade9 100644
--- a/arch/arm/cpu/Kconfig
+++ b/arch/arm/cpu/Kconfig
@@ -82,6 +82,7 @@ config CPU_V7
 	bool "Generic ARMv7 support"
 	depends on 32BIT
 	select CPU_32v7
+	select HAS_PBL_CLOCKSOURCE if !CPU_32v4T && !CPU_32v5 && !CPU_32v6
 
 config CPU_V7VE
 	bool "ARMv7 Virtualization Extensions Support"
@@ -135,6 +136,7 @@ config CPU_32v7
 config CPU_64v8
 	bool
 	select CPU_64
+	select HAS_PBL_CLOCKSOURCE
 
 comment "processor features"
 
diff --git a/arch/arm/lib32/Makefile b/arch/arm/lib32/Makefile
index 67c4f16f76..796ea6c742 100644
--- a/arch/arm/lib32/Makefile
+++ b/arch/arm/lib32/Makefile
@@ -34,7 +34,7 @@ extra-y += barebox.lds
 pbl-y	+= lib1funcs.o
 pbl-y	+= ashldi3.o
 pbl-y	+= div0.o
-pbl-$(CONFIG_CPU_32v7)	+= arm_architected_timer.o
+pbl-$(CONFIG_HAS_PBL_CLOCKSOURCE)	+= arm_architected_timer.o
 CFLAGS_arm_architected_timer.o := -march=armv7-a
 
 obj-pbl-y	+= setjmp.o
diff --git a/include/linux/iopoll.h b/include/linux/iopoll.h
index a6fade2a11..d86069f348 100644
--- a/include/linux/iopoll.h
+++ b/include/linux/iopoll.h
@@ -12,7 +12,7 @@
 #include <clock.h>
 #include <pbl.h>
 
-#if IN_PROPER
+#if IN_PROPER || IS_ENABLED(CONFIG_HAS_PBL_CLOCKSOURCE)
 # define read_poll_get_time_ns()	get_time_ns()
 # define read_poll_is_timeout(s, t)	is_timeout(s, t)
 #else
diff --git a/pbl/Kconfig b/pbl/Kconfig
index 63f29cd613..e5fe5328d2 100644
--- a/pbl/Kconfig
+++ b/pbl/Kconfig
@@ -43,6 +43,9 @@ config PBL_SINGLE_IMAGE
 	depends on !HAVE_PBL_MULTI_IMAGES
 	default y
 
+config HAS_PBL_CLOCKSOURCE
+	bool
+
 if PBL_IMAGE
 
 config USE_COMPRESSED_DTB

-- 
2.47.3

Re: [PATCH v2 2/3] ARM: lib32: pbl: implement get_time_ns and is_timeout

[Ahmad Fatoum]

Hi,

On 8/11/26 12:13 PM, Stefan Kerkmann wrote:
> ARMv7-A cores can implement the optional 64bit wide generic timer
> CNTPCT[1]. If it is present we can use it to implement the get_time_ns
> and is_timeout helpers which in turn enable the whole read_poll_timeout
> class of functions in the PBL. As it is guaranteed that the timer won't
> wrap for 40 years no overflow handling is necessary.
> 
> [1]: See "ARM Architecture Reference Manual ARMv7-A and ARMv7-R edition
> (rev C.d)", Chapter B8 "The Generic Timer"
> 
> Signed-off-by: Stefan Kerkmann <s.kerkmann@pengutronix.de>
> ---
>  arch/arm/lib32/arm_architected_timer.c | 15 ++++++++++++++-
>  1 file changed, 14 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm/lib32/arm_architected_timer.c b/arch/arm/lib32/arm_architected_timer.c
> index 54eca13f8b..75f74cbc15 100644
> --- a/arch/arm/lib32/arm_architected_timer.c
> +++ b/arch/arm/lib32/arm_architected_timer.c
> @@ -10,8 +10,21 @@ void arm_architected_timer_udelay(unsigned long us)
>  	unsigned long long ticks, cntfrq = get_cntfrq();
>  	unsigned long long start = get_cntpct();
>  
> -	ticks = DIV_ROUND_DOWN_ULL((us * cntfrq), 1000000);
> +	ticks = DIV_ROUND_DOWN_ULL((us * cntfrq), MSECOND);
>  
>  	while ((long)(start + ticks - get_cntpct()) > 0)
>  		;
>  }
> +
> +uint64_t get_time_ns(void)
> +{
> +	return get_cntpct() * SECOND / get_cntfrq();

I think you need to use one of the 64-bit division macros here.

> +int is_timeout(uint64_t start, uint64_t time_offset_ns)
> +{
> +	if ((int64_t)(start + time_offset_ns - get_time_ns()) < 0)
> +		return 1;
> +	else
> +		return 0;

That's a convoluted way to write

  return get_time_ns() >= start + time_offset_ns;

Why did you structure it this way with the cast?

(A 64-bit nanosecond timestamp overflows after hundreds of years).

Cheers,
Ahmad

> +}
> 

-- 
Pengutronix e.K.                  |                             |
Steuerwalder Str. 21              | http://www.pengutronix.de/  |
31137 Hildesheim, Germany         | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686  | Fax:   +49-5121-206917-5555 |

Re: [PATCH v2 3/3] pbl: introduce HAS_PBL_CLOCKSOURCE marker

[Sascha Hauer]

On 2026-08-11 12:13, Stefan Kerkmann wrote:
> This finally enables the usage of the polled timeout functions in the
> barebox pbl with real timeouts for supported architectures. Currently
> only ARMv7 and AARCH64 are enabled.
> 
> Signed-off-by: Stefan Kerkmann <s.kerkmann@pengutronix.de>
> ---
>  arch/arm/cpu/Kconfig    | 2 ++
>  arch/arm/lib32/Makefile | 2 +-
>  include/linux/iopoll.h  | 2 +-
>  pbl/Kconfig             | 3 +++
>  4 files changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/arm/cpu/Kconfig b/arch/arm/cpu/Kconfig
> index 47c116ea98..684c7eade9 100644
> --- a/arch/arm/cpu/Kconfig
> +++ b/arch/arm/cpu/Kconfig
> @@ -82,6 +82,7 @@ config CPU_V7
>  	bool "Generic ARMv7 support"
>  	depends on 32BIT
>  	select CPU_32v7
> +	select HAS_PBL_CLOCKSOURCE if !CPU_32v4T && !CPU_32v5 && !CPU_32v6
>  
>  config CPU_V7VE
>  	bool "ARMv7 Virtualization Extensions Support"
> @@ -135,6 +136,7 @@ config CPU_32v7
>  config CPU_64v8
>  	bool
>  	select CPU_64
> +	select HAS_PBL_CLOCKSOURCE
>  
>  comment "processor features"
>  
> diff --git a/arch/arm/lib32/Makefile b/arch/arm/lib32/Makefile
> index 67c4f16f76..796ea6c742 100644
> --- a/arch/arm/lib32/Makefile
> +++ b/arch/arm/lib32/Makefile
> @@ -34,7 +34,7 @@ extra-y += barebox.lds
>  pbl-y	+= lib1funcs.o
>  pbl-y	+= ashldi3.o
>  pbl-y	+= div0.o
> -pbl-$(CONFIG_CPU_32v7)	+= arm_architected_timer.o
> +pbl-$(CONFIG_HAS_PBL_CLOCKSOURCE)	+= arm_architected_timer.o

This change makes arm_architected_timer_udelay() available for builds
which exclusively support ARMv7a, so breaks builds that have ARMv4/5/6
enabled as well.

You could leave pbl-$(CONFIG_CPU_32v7) untouched and put get_time_ns()
inside #ifdef CONFIG_HAS_PBL_CLOCKSOURCE

Sascha

--
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |

Re: [PATCH v2 2/3] ARM: lib32: pbl: implement get_time_ns and is_timeout

[Sascha Hauer]

On 2026-08-11 12:21, Ahmad Fatoum wrote:
> Hi,
> 
> On 8/11/26 12:13 PM, Stefan Kerkmann wrote:
> > ARMv7-A cores can implement the optional 64bit wide generic timer
> > CNTPCT[1]. If it is present we can use it to implement the get_time_ns
> > and is_timeout helpers which in turn enable the whole read_poll_timeout
> > class of functions in the PBL. As it is guaranteed that the timer won't
> > wrap for 40 years no overflow handling is necessary.
> > 
> > [1]: See "ARM Architecture Reference Manual ARMv7-A and ARMv7-R edition
> > (rev C.d)", Chapter B8 "The Generic Timer"
> > 
> > Signed-off-by: Stefan Kerkmann <s.kerkmann@pengutronix.de>
> > ---
> >  arch/arm/lib32/arm_architected_timer.c | 15 ++++++++++++++-
> >  1 file changed, 14 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm/lib32/arm_architected_timer.c b/arch/arm/lib32/arm_architected_timer.c
> > index 54eca13f8b..75f74cbc15 100644
> > --- a/arch/arm/lib32/arm_architected_timer.c
> > +++ b/arch/arm/lib32/arm_architected_timer.c
> > @@ -10,8 +10,21 @@ void arm_architected_timer_udelay(unsigned long us)
> >  	unsigned long long ticks, cntfrq = get_cntfrq();
> >  	unsigned long long start = get_cntpct();
> >  
> > -	ticks = DIV_ROUND_DOWN_ULL((us * cntfrq), 1000000);
> > +	ticks = DIV_ROUND_DOWN_ULL((us * cntfrq), MSECOND);
> >  
> >  	while ((long)(start + ticks - get_cntpct()) > 0)
> >  		;
> >  }
> > +
> > +uint64_t get_time_ns(void)
> > +{
> > +	return get_cntpct() * SECOND / get_cntfrq();
> 
> I think you need to use one of the 64-bit division macros here.

Also the corresponding barebox proper code doesn't depend on the timer
being initialized to 0. I don't know if we can guarantee this for the
architected timer, but otherwise get_cntpct() * SECOND could easily
overflow.

> 
> > +int is_timeout(uint64_t start, uint64_t time_offset_ns)
> > +{
> > +	if ((int64_t)(start + time_offset_ns - get_time_ns()) < 0)
> > +		return 1;
> > +	else
> > +		return 0;
> 
> That's a convoluted way to write
> 
>   return get_time_ns() >= start + time_offset_ns;
> 
> Why did you structure it this way with the cast?
> 
> (A 64-bit nanosecond timestamp overflows after hundreds of years).

It's a direct copy of the non-pbl is_timeout implementation. It's
agnostic to the timer initialization value which is not necessarily 0.

Sascha

--
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |