[PATCHv4,1/2] ARM: imx53: add secure-reg-access support for PMU

Message ID 20180212123945.15732-2-sebastian.reichel@collabora.co.uk
State New
Headers show
Series
  • Improved perf support for imx53/ppd
Related show

Commit Message

Sebastian Reichel Feb. 12, 2018, 12:39 p.m.
On i.MX53 it is necessary to set the DBG_EN bit in the
platform GPC register to enable access to PMU counters
other than the cycle counter.

Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>
---
 arch/arm/mach-imx/mach-imx53.c | 39 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 38 insertions(+), 1 deletion(-)

Comments

Fabio Estevam Feb. 12, 2018, 12:48 p.m. | #1
On Mon, Feb 12, 2018 at 10:39 AM, Sebastian Reichel
<sebastian.reichel@collabora.co.uk> wrote:
> On i.MX53 it is necessary to set the DBG_EN bit in the
> platform GPC register to enable access to PMU counters
> other than the cycle counter.
>
> Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>

Reviewed-by: Fabio Estevam <fabio.estevam@nxp.com>
Shawn Guo Feb. 24, 2018, 7:45 a.m. | #2
On Mon, Feb 12, 2018 at 01:39:44PM +0100, Sebastian Reichel wrote:
> On i.MX53 it is necessary to set the DBG_EN bit in the
> platform GPC register to enable access to PMU counters
> other than the cycle counter.
> 
> Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>
> ---
>  arch/arm/mach-imx/mach-imx53.c | 39 ++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 38 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm/mach-imx/mach-imx53.c b/arch/arm/mach-imx/mach-imx53.c
> index 07c2e8dca494..658e28604dca 100644
> --- a/arch/arm/mach-imx/mach-imx53.c
> +++ b/arch/arm/mach-imx/mach-imx53.c
> @@ -28,10 +28,47 @@ static void __init imx53_init_early(void)
>  	mxc_set_cpu_type(MXC_CPU_MX53);
>  }
>  
> +#define MXC_CORTEXA8_PLAT_GPC 0x63fa0004

The base address should be retrieved from device tree.

Shawn

> +#define GPC_DBG_EN BIT(16)
> +
> +/*
> + * This enables the DBGEN bit in ARM_GPC register, which is
> + * required for accessing some performance counter features.
> + * Technically it is only required while perf is used, but to
> + * keep the source code simple we just enable it all the time
> + * when the kernel configuration allows using the feature.
> + */
> +static void imx53_pmu_init(void)
> +{
> +	void __iomem *gpc_reg;
> +	struct device_node *node;
> +	u32 gpc;
> +
> +	if (!IS_ENABLED(CONFIG_ARM_PMU))
> +		return;
> +
> +	node = of_find_compatible_node(NULL, NULL, "arm,cortex-a8-pmu");
> +	if (!node)
> +		return;
> +
> +	if (!of_property_read_bool(node, "secure-reg-access"))
> +		return;
> +
> +	gpc_reg = ioremap(MXC_CORTEXA8_PLAT_GPC, 4);
> +	if (!gpc_reg) {
> +		pr_warning("unable to map GPC to enable perf\n");
> +		return;
> +	}
> +
> +	gpc = readl_relaxed(gpc_reg);
> +	gpc |= GPC_DBG_EN;
> +	writel_relaxed(gpc, gpc_reg);
> +}
> +
>  static void __init imx53_dt_init(void)
>  {
>  	imx_src_init();
> -
> +	imx53_pmu_init();
>  	imx_aips_allow_unprivileged_access("fsl,imx53-aipstz");
>  }
>  
> -- 
> 2.15.1
>
Sebastian Reichel Feb. 26, 2018, 1:47 p.m. | #3
Hi Shawn,

On Sat, Feb 24, 2018 at 03:45:44PM +0800, Shawn Guo wrote:
> On Mon, Feb 12, 2018 at 01:39:44PM +0100, Sebastian Reichel wrote:
> > On i.MX53 it is necessary to set the DBG_EN bit in the
> > platform GPC register to enable access to PMU counters
> > other than the cycle counter.
> > 
> > Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>
> > ---
> >  arch/arm/mach-imx/mach-imx53.c | 39 ++++++++++++++++++++++++++++++++++++++-
> >  1 file changed, 38 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm/mach-imx/mach-imx53.c b/arch/arm/mach-imx/mach-imx53.c
> > index 07c2e8dca494..658e28604dca 100644
> > --- a/arch/arm/mach-imx/mach-imx53.c
> > +++ b/arch/arm/mach-imx/mach-imx53.c
> > @@ -28,10 +28,47 @@ static void __init imx53_init_early(void)
> >  	mxc_set_cpu_type(MXC_CPU_MX53);
> >  }
> >  
> > +#define MXC_CORTEXA8_PLAT_GPC 0x63fa0004
> 
> The base address should be retrieved from device tree.

DT has no entry for 0x63fa0000-0x63fa3fff. iMX53 TRM lists it as "ARM Platform"
with 8 platform specific 32 bit registers. Do you think it's worth the trouble
adding a new binding? Do you have a suggestion for a compatible value?

-- Sebastian

> 
> Shawn
> 
> > +#define GPC_DBG_EN BIT(16)
> > +
> > +/*
> > + * This enables the DBGEN bit in ARM_GPC register, which is
> > + * required for accessing some performance counter features.
> > + * Technically it is only required while perf is used, but to
> > + * keep the source code simple we just enable it all the time
> > + * when the kernel configuration allows using the feature.
> > + */
> > +static void imx53_pmu_init(void)
> > +{
> > +	void __iomem *gpc_reg;
> > +	struct device_node *node;
> > +	u32 gpc;
> > +
> > +	if (!IS_ENABLED(CONFIG_ARM_PMU))
> > +		return;
> > +
> > +	node = of_find_compatible_node(NULL, NULL, "arm,cortex-a8-pmu");
> > +	if (!node)
> > +		return;
> > +
> > +	if (!of_property_read_bool(node, "secure-reg-access"))
> > +		return;
> > +
> > +	gpc_reg = ioremap(MXC_CORTEXA8_PLAT_GPC, 4);
> > +	if (!gpc_reg) {
> > +		pr_warning("unable to map GPC to enable perf\n");
> > +		return;
> > +	}
> > +
> > +	gpc = readl_relaxed(gpc_reg);
> > +	gpc |= GPC_DBG_EN;
> > +	writel_relaxed(gpc, gpc_reg);
> > +}
> > +
> >  static void __init imx53_dt_init(void)
> >  {
> >  	imx_src_init();
> > -
> > +	imx53_pmu_init();
> >  	imx_aips_allow_unprivileged_access("fsl,imx53-aipstz");
> >  }
> >  
> > -- 
> > 2.15.1
> >
Shawn Guo Feb. 27, 2018, 1:10 a.m. | #4
On Mon, Feb 26, 2018 at 02:47:41PM +0100, Sebastian Reichel wrote:
> Hi Shawn,
> 
> On Sat, Feb 24, 2018 at 03:45:44PM +0800, Shawn Guo wrote:
> > On Mon, Feb 12, 2018 at 01:39:44PM +0100, Sebastian Reichel wrote:
> > > On i.MX53 it is necessary to set the DBG_EN bit in the
> > > platform GPC register to enable access to PMU counters
> > > other than the cycle counter.
> > > 
> > > Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>
> > > ---
> > >  arch/arm/mach-imx/mach-imx53.c | 39 ++++++++++++++++++++++++++++++++++++++-
> > >  1 file changed, 38 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/arch/arm/mach-imx/mach-imx53.c b/arch/arm/mach-imx/mach-imx53.c
> > > index 07c2e8dca494..658e28604dca 100644
> > > --- a/arch/arm/mach-imx/mach-imx53.c
> > > +++ b/arch/arm/mach-imx/mach-imx53.c
> > > @@ -28,10 +28,47 @@ static void __init imx53_init_early(void)
> > >  	mxc_set_cpu_type(MXC_CPU_MX53);
> > >  }
> > >  
> > > +#define MXC_CORTEXA8_PLAT_GPC 0x63fa0004
> > 
> > The base address should be retrieved from device tree.
> 
> DT has no entry for 0x63fa0000-0x63fa3fff. iMX53 TRM lists it as "ARM Platform"
> with 8 platform specific 32 bit registers. Do you think it's worth the trouble
> adding a new binding? Do you have a suggestion for a compatible value?

Looking at it more closely, I feel that patching every single platform
which needs to set up additional register for secure-reg-access support
doesn't really scale.  Can we have pmu driver do it with a phandle in
DT pointing to the register and bit that need to be configured?

Shawn

> > > +#define GPC_DBG_EN BIT(16)
> > > +
> > > +/*
> > > + * This enables the DBGEN bit in ARM_GPC register, which is
> > > + * required for accessing some performance counter features.
> > > + * Technically it is only required while perf is used, but to
> > > + * keep the source code simple we just enable it all the time
> > > + * when the kernel configuration allows using the feature.
> > > + */
> > > +static void imx53_pmu_init(void)
> > > +{
> > > +	void __iomem *gpc_reg;
> > > +	struct device_node *node;
> > > +	u32 gpc;
> > > +
> > > +	if (!IS_ENABLED(CONFIG_ARM_PMU))
> > > +		return;
> > > +
> > > +	node = of_find_compatible_node(NULL, NULL, "arm,cortex-a8-pmu");
> > > +	if (!node)
> > > +		return;
> > > +
> > > +	if (!of_property_read_bool(node, "secure-reg-access"))
> > > +		return;
> > > +
> > > +	gpc_reg = ioremap(MXC_CORTEXA8_PLAT_GPC, 4);
> > > +	if (!gpc_reg) {
> > > +		pr_warning("unable to map GPC to enable perf\n");
> > > +		return;
> > > +	}
> > > +
> > > +	gpc = readl_relaxed(gpc_reg);
> > > +	gpc |= GPC_DBG_EN;
> > > +	writel_relaxed(gpc, gpc_reg);
> > > +}
> > > +
> > >  static void __init imx53_dt_init(void)
> > >  {
> > >  	imx_src_init();
> > > -
> > > +	imx53_pmu_init();
> > >  	imx_aips_allow_unprivileged_access("fsl,imx53-aipstz");
> > >  }
> > >  
> > > -- 
> > > 2.15.1
> > >
Sebastian Reichel Feb. 27, 2018, 10:17 a.m. | #5
Hi,

On Tue, Feb 27, 2018 at 09:10:34AM +0800, Shawn Guo wrote:
> On Mon, Feb 26, 2018 at 02:47:41PM +0100, Sebastian Reichel wrote:
> > On Sat, Feb 24, 2018 at 03:45:44PM +0800, Shawn Guo wrote:
> > > On Mon, Feb 12, 2018 at 01:39:44PM +0100, Sebastian Reichel wrote:
> > > > On i.MX53 it is necessary to set the DBG_EN bit in the
> > > > platform GPC register to enable access to PMU counters
> > > > other than the cycle counter.
> > > > 
> > > > Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>
> > > > ---
> > > >  arch/arm/mach-imx/mach-imx53.c | 39 ++++++++++++++++++++++++++++++++++++++-
> > > >  1 file changed, 38 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/arch/arm/mach-imx/mach-imx53.c b/arch/arm/mach-imx/mach-imx53.c
> > > > index 07c2e8dca494..658e28604dca 100644
> > > > --- a/arch/arm/mach-imx/mach-imx53.c
> > > > +++ b/arch/arm/mach-imx/mach-imx53.c
> > > > @@ -28,10 +28,47 @@ static void __init imx53_init_early(void)
> > > >  	mxc_set_cpu_type(MXC_CPU_MX53);
> > > >  }
> > > >  
> > > > +#define MXC_CORTEXA8_PLAT_GPC 0x63fa0004
> > > 
> > > The base address should be retrieved from device tree.
> > 
> > DT has no entry for 0x63fa0000-0x63fa3fff. iMX53 TRM lists it as "ARM Platform"
> > with 8 platform specific 32 bit registers. Do you think it's worth the trouble
> > adding a new binding? Do you have a suggestion for a compatible value?
> 
> Looking at it more closely, I feel that patching every single platform
> which needs to set up additional register for secure-reg-access support
> doesn't really scale.  Can we have pmu driver do it with a phandle in
> DT pointing to the register and bit that need to be configured?

The PMU driver used to have a feature for initialising platform
specific bits, but it is currently being removed to make the PMU
code more maintainable. My understanding is, that it's very uncommon
to require platform specific setup to get secure-reg-access working.
I.e. it is not needed for newer iMX platforms.

-- Sebastian

> 
> Shawn
> 
> > > > +#define GPC_DBG_EN BIT(16)
> > > > +
> > > > +/*
> > > > + * This enables the DBGEN bit in ARM_GPC register, which is
> > > > + * required for accessing some performance counter features.
> > > > + * Technically it is only required while perf is used, but to
> > > > + * keep the source code simple we just enable it all the time
> > > > + * when the kernel configuration allows using the feature.
> > > > + */
> > > > +static void imx53_pmu_init(void)
> > > > +{
> > > > +	void __iomem *gpc_reg;
> > > > +	struct device_node *node;
> > > > +	u32 gpc;
> > > > +
> > > > +	if (!IS_ENABLED(CONFIG_ARM_PMU))
> > > > +		return;
> > > > +
> > > > +	node = of_find_compatible_node(NULL, NULL, "arm,cortex-a8-pmu");
> > > > +	if (!node)
> > > > +		return;
> > > > +
> > > > +	if (!of_property_read_bool(node, "secure-reg-access"))
> > > > +		return;
> > > > +
> > > > +	gpc_reg = ioremap(MXC_CORTEXA8_PLAT_GPC, 4);
> > > > +	if (!gpc_reg) {
> > > > +		pr_warning("unable to map GPC to enable perf\n");
> > > > +		return;
> > > > +	}
> > > > +
> > > > +	gpc = readl_relaxed(gpc_reg);
> > > > +	gpc |= GPC_DBG_EN;
> > > > +	writel_relaxed(gpc, gpc_reg);
> > > > +}
> > > > +
> > > >  static void __init imx53_dt_init(void)
> > > >  {
> > > >  	imx_src_init();
> > > > -
> > > > +	imx53_pmu_init();
> > > >  	imx_aips_allow_unprivileged_access("fsl,imx53-aipstz");
> > > >  }
> > > >  
> > > > -- 
> > > > 2.15.1
> > > > 
> 
>
Sebastian Reichel May 25, 2018, 3:45 p.m. | #6
Hi Shawn,

On Tue, Feb 27, 2018 at 11:17:12AM +0100, Sebastian Reichel wrote:
> Hi,
> 
> On Tue, Feb 27, 2018 at 09:10:34AM +0800, Shawn Guo wrote:
> > On Mon, Feb 26, 2018 at 02:47:41PM +0100, Sebastian Reichel wrote:
> > > On Sat, Feb 24, 2018 at 03:45:44PM +0800, Shawn Guo wrote:
> > > > On Mon, Feb 12, 2018 at 01:39:44PM +0100, Sebastian Reichel wrote:
> > > > > On i.MX53 it is necessary to set the DBG_EN bit in the
> > > > > platform GPC register to enable access to PMU counters
> > > > > other than the cycle counter.
> > > > > 
> > > > > Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>
> > > > > ---
> > > > >  arch/arm/mach-imx/mach-imx53.c | 39 ++++++++++++++++++++++++++++++++++++++-
> > > > >  1 file changed, 38 insertions(+), 1 deletion(-)
> > > > > 
> > > > > diff --git a/arch/arm/mach-imx/mach-imx53.c b/arch/arm/mach-imx/mach-imx53.c
> > > > > index 07c2e8dca494..658e28604dca 100644
> > > > > --- a/arch/arm/mach-imx/mach-imx53.c
> > > > > +++ b/arch/arm/mach-imx/mach-imx53.c
> > > > > @@ -28,10 +28,47 @@ static void __init imx53_init_early(void)
> > > > >  	mxc_set_cpu_type(MXC_CPU_MX53);
> > > > >  }
> > > > >  
> > > > > +#define MXC_CORTEXA8_PLAT_GPC 0x63fa0004
> > > > 
> > > > The base address should be retrieved from device tree.
> > > 
> > > DT has no entry for 0x63fa0000-0x63fa3fff. iMX53 TRM lists it as "ARM Platform"
> > > with 8 platform specific 32 bit registers. Do you think it's worth the trouble
> > > adding a new binding? Do you have a suggestion for a compatible value?
> > 
> > Looking at it more closely, I feel that patching every single platform
> > which needs to set up additional register for secure-reg-access support
> > doesn't really scale.  Can we have pmu driver do it with a phandle in
> > DT pointing to the register and bit that need to be configured?
> 
> The PMU driver used to have a feature for initialising platform
> specific bits, but it is currently being removed to make the PMU
> code more maintainable. My understanding is, that it's very uncommon
> to require platform specific setup to get secure-reg-access working.
> E.g. it is not needed for newer iMX platforms.

You never replied to this one. Are you fine with adding the
function? This is kind of a deadlock situation with neither
you wanting to enable the platform bit, nor the generic PMU
driver.

-- Sebastian
Shawn Guo May 28, 2018, 2:26 a.m. | #7
On Tue, Feb 27, 2018 at 11:17:12AM +0100, Sebastian Reichel wrote:
> Hi,
> 
> On Tue, Feb 27, 2018 at 09:10:34AM +0800, Shawn Guo wrote:
> > On Mon, Feb 26, 2018 at 02:47:41PM +0100, Sebastian Reichel wrote:
> > > On Sat, Feb 24, 2018 at 03:45:44PM +0800, Shawn Guo wrote:
> > > > On Mon, Feb 12, 2018 at 01:39:44PM +0100, Sebastian Reichel wrote:
> > > > > On i.MX53 it is necessary to set the DBG_EN bit in the
> > > > > platform GPC register to enable access to PMU counters
> > > > > other than the cycle counter.
> > > > > 
> > > > > Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>
> > > > > ---
> > > > >  arch/arm/mach-imx/mach-imx53.c | 39 ++++++++++++++++++++++++++++++++++++++-
> > > > >  1 file changed, 38 insertions(+), 1 deletion(-)
> > > > > 
> > > > > diff --git a/arch/arm/mach-imx/mach-imx53.c b/arch/arm/mach-imx/mach-imx53.c
> > > > > index 07c2e8dca494..658e28604dca 100644
> > > > > --- a/arch/arm/mach-imx/mach-imx53.c
> > > > > +++ b/arch/arm/mach-imx/mach-imx53.c
> > > > > @@ -28,10 +28,47 @@ static void __init imx53_init_early(void)
> > > > >  	mxc_set_cpu_type(MXC_CPU_MX53);
> > > > >  }
> > > > >  
> > > > > +#define MXC_CORTEXA8_PLAT_GPC 0x63fa0004
> > > > 
> > > > The base address should be retrieved from device tree.
> > > 
> > > DT has no entry for 0x63fa0000-0x63fa3fff. iMX53 TRM lists it as "ARM Platform"
> > > with 8 platform specific 32 bit registers. Do you think it's worth the trouble
> > > adding a new binding? Do you have a suggestion for a compatible value?
> > 
> > Looking at it more closely, I feel that patching every single platform
> > which needs to set up additional register for secure-reg-access support
> > doesn't really scale.  Can we have pmu driver do it with a phandle in
> > DT pointing to the register and bit that need to be configured?
> 
> The PMU driver used to have a feature for initialising platform
> specific bits, but it is currently being removed to make the PMU
> code more maintainable. My understanding is, that it's very uncommon
> to require platform specific setup to get secure-reg-access working.
> I.e. it is not needed for newer iMX platforms.

Are you saying this is a very specific setup required by i.MX53 only?
In that case, I can live with it.

Shawn
Sebastian Reichel May 28, 2018, 6:41 a.m. | #8
Hi,

On Mon, May 28, 2018 at 10:26:54AM +0800, Shawn Guo wrote:
> On Tue, Feb 27, 2018 at 11:17:12AM +0100, Sebastian Reichel wrote:
> > Hi,
> > 
> > On Tue, Feb 27, 2018 at 09:10:34AM +0800, Shawn Guo wrote:
> > > On Mon, Feb 26, 2018 at 02:47:41PM +0100, Sebastian Reichel wrote:
> > > > On Sat, Feb 24, 2018 at 03:45:44PM +0800, Shawn Guo wrote:
> > > > > On Mon, Feb 12, 2018 at 01:39:44PM +0100, Sebastian Reichel wrote:
> > > > > > On i.MX53 it is necessary to set the DBG_EN bit in the
> > > > > > platform GPC register to enable access to PMU counters
> > > > > > other than the cycle counter.
> > > > > > 
> > > > > > Signed-off-by: Sebastian Reichel <sebastian.reichel@collabora.co.uk>
> > > > > > ---
> > > > > >  arch/arm/mach-imx/mach-imx53.c | 39 ++++++++++++++++++++++++++++++++++++++-
> > > > > >  1 file changed, 38 insertions(+), 1 deletion(-)
> > > > > > 
> > > > > > diff --git a/arch/arm/mach-imx/mach-imx53.c b/arch/arm/mach-imx/mach-imx53.c
> > > > > > index 07c2e8dca494..658e28604dca 100644
> > > > > > --- a/arch/arm/mach-imx/mach-imx53.c
> > > > > > +++ b/arch/arm/mach-imx/mach-imx53.c
> > > > > > @@ -28,10 +28,47 @@ static void __init imx53_init_early(void)
> > > > > >  	mxc_set_cpu_type(MXC_CPU_MX53);
> > > > > >  }
> > > > > >  
> > > > > > +#define MXC_CORTEXA8_PLAT_GPC 0x63fa0004
> > > > > 
> > > > > The base address should be retrieved from device tree.
> > > > 
> > > > DT has no entry for 0x63fa0000-0x63fa3fff. iMX53 TRM lists it as "ARM Platform"
> > > > with 8 platform specific 32 bit registers. Do you think it's worth the trouble
> > > > adding a new binding? Do you have a suggestion for a compatible value?
> > > 
> > > Looking at it more closely, I feel that patching every single platform
> > > which needs to set up additional register for secure-reg-access support
> > > doesn't really scale.  Can we have pmu driver do it with a phandle in
> > > DT pointing to the register and bit that need to be configured?
> > 
> > The PMU driver used to have a feature for initialising platform
> > specific bits, but it is currently being removed to make the PMU
> > code more maintainable. My understanding is, that it's very uncommon
> > to require platform specific setup to get secure-reg-access working.
> > I.e. it is not needed for newer iMX platforms.
> 
> Are you saying this is a very specific setup required by i.MX53 only?

Yes, all other SoCs supported by Linux ARM PMU counters driver can
just use the registers without having to enable platform specific
bits first.

> In that case, I can live with it.

What about the DT node? I did not add it, since this is a i.MX53
specific workaround anyways.

-- Sebastian
Shawn Guo May 28, 2018, 7:20 a.m. | #9
On Mon, May 28, 2018 at 08:41:31AM +0200, Sebastian Reichel wrote:
> > Are you saying this is a very specific setup required by i.MX53 only?
> 
> Yes, all other SoCs supported by Linux ARM PMU counters driver can
> just use the registers without having to enable platform specific
> bits first.
> 
> > In that case, I can live with it.
> 
> What about the DT node? I did not add it, since this is a i.MX53
> specific workaround anyways.

What you are adding here is secure-reg-access property, which has an
defined meaning in PMU binding doc.  I'm not really sure if it's
appropriate to use the property as a condition for DBGEN bit setup.
Or can we set up the bit regardless of the property?

Shawn
Sebastian Reichel May 28, 2018, 3:50 p.m. | #10
Hi,

On Mon, May 28, 2018 at 03:20:35PM +0800, Shawn Guo wrote:
> On Mon, May 28, 2018 at 08:41:31AM +0200, Sebastian Reichel wrote:
> > > Are you saying this is a very specific setup required by i.MX53 only?
> > 
> > Yes, all other SoCs supported by Linux ARM PMU counters driver can
> > just use the registers without having to enable platform specific
> > bits first.
> > 
> > > In that case, I can live with it.
> > 
> > What about the DT node? I did not add it, since this is a i.MX53
> > specific workaround anyways.
> 
> What you are adding here is secure-reg-access property, which has an
> defined meaning in PMU binding doc.  I'm not really sure if it's
> appropriate to use the property as a condition for DBGEN bit setup.
> Or can we set up the bit regardless of the property?

The description for DBGEN bit is:

> Debug enable. This allows the user to manually activate clocks
> within the debug system. This register bit directly controls the
> platform's dbgen_out output signal which connects to the DAP_SYS to
> enable all debug clocks. Once enabled, the clocks cannot be disabled
> except by asserting the disable_trace input of the DAP_SYS.

I only enable this bit when the kernel configuration allows
using the PMU, since otherwise we do not need the clocks
in the debug system. This limits any potential side-effects
of this patchset.

-- Sebastian

Patch

diff --git a/arch/arm/mach-imx/mach-imx53.c b/arch/arm/mach-imx/mach-imx53.c
index 07c2e8dca494..658e28604dca 100644
--- a/arch/arm/mach-imx/mach-imx53.c
+++ b/arch/arm/mach-imx/mach-imx53.c
@@ -28,10 +28,47 @@  static void __init imx53_init_early(void)
 	mxc_set_cpu_type(MXC_CPU_MX53);
 }
 
+#define MXC_CORTEXA8_PLAT_GPC 0x63fa0004
+#define GPC_DBG_EN BIT(16)
+
+/*
+ * This enables the DBGEN bit in ARM_GPC register, which is
+ * required for accessing some performance counter features.
+ * Technically it is only required while perf is used, but to
+ * keep the source code simple we just enable it all the time
+ * when the kernel configuration allows using the feature.
+ */
+static void imx53_pmu_init(void)
+{
+	void __iomem *gpc_reg;
+	struct device_node *node;
+	u32 gpc;
+
+	if (!IS_ENABLED(CONFIG_ARM_PMU))
+		return;
+
+	node = of_find_compatible_node(NULL, NULL, "arm,cortex-a8-pmu");
+	if (!node)
+		return;
+
+	if (!of_property_read_bool(node, "secure-reg-access"))
+		return;
+
+	gpc_reg = ioremap(MXC_CORTEXA8_PLAT_GPC, 4);
+	if (!gpc_reg) {
+		pr_warning("unable to map GPC to enable perf\n");
+		return;
+	}
+
+	gpc = readl_relaxed(gpc_reg);
+	gpc |= GPC_DBG_EN;
+	writel_relaxed(gpc, gpc_reg);
+}
+
 static void __init imx53_dt_init(void)
 {
 	imx_src_init();
-
+	imx53_pmu_init();
 	imx_aips_allow_unprivileged_access("fsl,imx53-aipstz");
 }