Skip to contentSkip to conversation

ARM: mach-shmobile: Don't configure ARCH timer if PSCI is enabled

PATCH In review Oleksandr Tyshchenko 3 participants 7 messages Apr 17, 2019 – Apr 26, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ARM: mach-shmobile: Don't configure ARCH timer if PSCI is enabled

From: Oleksandr Tyshchenko <[email protected]>

If CONFIG_PSCI is enabled then most likely we are running on PSCI-enabled U-Boot which, we assume, has already taken care of configuring ARCH timer stuff before switching to non-secure mode.

---
A bit of context here...

We are highly interested in Renesas "Stout" board support (r8a7790) in Xen hypervisor. The reason is to have fully supported HW for performing "OSSTEST" (Xen automatic test system) on ARM32.

To reach that target we need a "generic way" for the secondary CPU cores bring up and switching them to non-secure hyp mode.
So, the PSCI as a generic well-known way to bring up CPUs, was chosen for that purpose.

You can find corresponding patches for U-Boot here:
http://u-boot.10912.n7.nabble.com/PATCH-0-3-PSCI-support-for-r8a7790-SoC-Lager-Stout-boards-td357352.html

You can find corresponding patches for Xen hypervisor here:
https://www.mail-archive.com/[email protected]/msg43332.html

To sumarize:
Together with enabling CONFIG_PSCI in shmobile_defconfig, current patch is a minimal required change needed to run mainline Linux on top of PSCI-enabled U-Boot.
There is no need to modify device tree. U-Boot will take care of inserting proper "enable-method" strings in CPU nodes.

---
 arch/arm/mach-shmobile/setup-rcar-gen2.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
arch/arm/mach-shmobile/setup-rcar-gen2.c +8 -1
index eea60b2..bac4490 100644
@@ -32,7 +32,7 @@ static const struct of_device_id cpg_matches[] __initconst = {
 	{ /* sentinel */ }
 };
 
-static unsigned int __init get_extal_freq(void)
+static unsigned int __init __maybe_unused get_extal_freq(void)
 {
 	const struct of_device_id *match;
 	struct device_node *cpg, *extal;
@@ -60,6 +60,12 @@ static unsigned int __init get_extal_freq(void)
 
 void __init rcar_gen2_timer_init(void)
 {
+/*
+ * If CONFIG_PSCI is enabled then most likely we are running on PSCI-enabled
+ * U-Boot which, we assume, has already taken care of configuring ARCH timer
+ * stuff before switching to non-secure mode.
+ */
+#if !defined(CONFIG_ARM_PSCI)
 	void __iomem *base;
 	u32 freq;
 
@@ -101,6 +107,7 @@ void __init rcar_gen2_timer_init(void)
 	}
 
 	iounmap(base);
+#endif /* #if !defined(CONFIG_ARM_PSCI) */
 
 	of_clk_init(NULL);
 	timer_probe();
Signature · 1 line
2.7.4
#1 lore raw

Hi Oleksandr,

Thanks for the patch.

Quoted 59 lines

Subject: [PATCH] ARM: mach-shmobile: Don't configure ARCH timer if PSCI is enabled

From: Oleksandr Tyshchenko <[email protected]>

If CONFIG_PSCI is enabled then most likely we are running on PSCI-enabled U-Boot which, we assume, has already taken care of configuring ARCH timer stuff before switching to non-secure mode.

---
A bit of context here...

We are highly interested in Renesas "Stout" board support (r8a7790) in Xen hypervisor. The reason is to have fully supported HW for performing "OSSTEST" (Xen automatic test system) on ARM32.

To reach that target we need a "generic way" for the secondary CPU cores bring up and switching them to non-secure hyp mode.
So, the PSCI as a generic well-known way to bring up CPUs, was chosen for that purpose.

You can find corresponding patches for U-Boot here:
http://u-boot.10912.n7.nabble.com/PATCH-0-3-PSCI-support-for-r8a7790-
SoC-Lager-Stout-boards-td357352.html

You can find corresponding patches for Xen hypervisor here:
https://www.mail-archive.com/xen-
[email protected]/msg43332.html

To sumarize:
Together with enabling CONFIG_PSCI in shmobile_defconfig, current patch is a minimal required change needed to run mainline Linux on top of PSCI- enabled U-Boot.
There is no need to modify device tree. U-Boot will take care of inserting proper "enable-method" strings in CPU nodes.

---
 arch/arm/mach-shmobile/setup-rcar-gen2.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
arch/arm/mach- renamed from arch/arm/mach-shmobile/setup-rcar-gen2.c +0 -0

shmobile/setup-rcar-gen2.c
index eea60b2..bac4490 100644

arch/arm/mach-shmobile/setup-rcar-gen2.c +1 -1
@@ -32,7 +32,7 @@ static const struct of_device_id cpg_matches[]
 __initconst = {
 	{ /* sentinel */ }
 };
 
-static unsigned int __init get_extal_freq(void)
+static unsigned int __init __maybe_unused get_extal_freq(void)
 {
 	const struct of_device_id *match;
arch/arm/mach-shmobile/setup-rcar-gen2.c +0 -0
 	struct device_node *cpg, *extal;
@@ -60,6 +60,12 @@ static unsigned int __init get_extal_freq(void)
 

arch/arm/mach-shmobile/setup-rcar-gen2.c static unsigned int __init get_extal_freq(void) · line 70

Quoted 10 lines
arch/arm/mach-shmobile/setup-rcar-gen2.c +8 -0
 void __init rcar_gen2_timer_init(void)
 {
+/*
+ * If CONFIG_PSCI is enabled then most likely we are running on
+PSCI-enabled
+ * U-Boot which, we assume, has already taken care of configuring ARCH
+timer
+ * stuff before switching to non-secure mode.
+ */
+#if !defined(CONFIG_ARM_PSCI)

Is it required? If you see the below comment, it is already taken care by the below code.

Is your code entering into this block, when booting the kernel in NS mode?

<snip>
/*

 87          * Update the timer if it is either not running, or is not at the        
 88          * right frequency. The timer is only configurable in secure mode        
 89          * so this avoids an abort if the loader started the timer and           
 90          * entered the kernel in non-secure mode.                                
 91          */                                                                      
 92                                                                                  
 93         if ((ioread32(base + CNTCR) & 1) == 0 ||                                 
 94             ioread32(base + CNTFID0) != freq) {                                  
 95                 /* Update registers with correct frequency */                    
 96                 iowrite32(freq, base + CNTFID0);                                 
 97                 asm volatile("mcr p15, 0, %0, c14, c0, 0" : : "r" (freq));       

98
99 /* make sure arch timer is started by setting bit 0 of CNTCR */
100 iowrite32(1, base + CNTCR);
101 }

Regards,
Biju

Quoted 13 lines
arch/arm/mach-shmobile/setup-rcar-gen2.c +1 -1
 	void __iomem *base;
 	u32 freq;
@@ -101,6 +107,7 @@ void __init rcar_gen2_timer_init(void)
 	}
 
 	iounmap(base);
+#endif /* #if !defined(CONFIG_ARM_PSCI) */
 
 	of_clk_init(NULL);
 	timer_probe();
--

2.7.4

author
#2 lore raw
Quoted 1 line · Biju Das

Hi Oleksandr,

Hi Biju

Quoted 75 lines

Thanks for the patch.

Quoted 59 lines

Subject: [PATCH] ARM: mach-shmobile: Don't configure ARCH timer if PSCI is enabled

From: Oleksandr Tyshchenko <[email protected]>

If CONFIG_PSCI is enabled then most likely we are running on PSCI-enabled U-Boot which, we assume, has already taken care of configuring ARCH timer stuff before switching to non-secure mode.

---
A bit of context here...

We are highly interested in Renesas "Stout" board support (r8a7790) in Xen hypervisor. The reason is to have fully supported HW for performing "OSSTEST" (Xen automatic test system) on ARM32.

To reach that target we need a "generic way" for the secondary CPU cores bring up and switching them to non-secure hyp mode.
So, the PSCI as a generic well-known way to bring up CPUs, was chosen for that purpose.

You can find corresponding patches for U-Boot here:
http://u-boot.10912.n7.nabble.com/PATCH-0-3-PSCI-support-for-r8a7790-
SoC-Lager-Stout-boards-td357352.html

You can find corresponding patches for Xen hypervisor here:
https://www.mail-archive.com/xen-
[email protected]/msg43332.html

To sumarize:
Together with enabling CONFIG_PSCI in shmobile_defconfig, current patch is a minimal required change needed to run mainline Linux on top of PSCI- enabled U-Boot.
There is no need to modify device tree. U-Boot will take care of inserting proper "enable-method" strings in CPU nodes.

---
  arch/arm/mach-shmobile/setup-rcar-gen2.c | 9 ++++++++-
  1 file changed, 8 insertions(+), 1 deletion(-)
arch/arm/mach- renamed from arch/arm/mach-shmobile/setup-rcar-gen2.c +0 -0

shmobile/setup-rcar-gen2.c
index eea60b2..bac4490 100644

arch/arm/mach-shmobile/setup-rcar-gen2.c +1 -1
@@ -32,7 +32,7 @@ static const struct of_device_id cpg_matches[]
 __initconst = {
  	{ /* sentinel */ }
  };
 
-static unsigned int __init get_extal_freq(void)
+static unsigned int __init __maybe_unused get_extal_freq(void)
  {
  	const struct of_device_id *match;

struct device_node *cpg, *extal;

arch/arm/mach-shmobile/setup-rcar-gen2.c +0 -0
@@ -60,6 +60,12 @@ static unsigned int __init get_extal_freq(void)
 

arch/arm/mach-shmobile/setup-rcar-gen2.c static unsigned int __init get_extal_freq(void) · line 70

Quoted 10 lines
arch/arm/mach-shmobile/setup-rcar-gen2.c +8 -0
  void __init rcar_gen2_timer_init(void)
  {
+/*
+ * If CONFIG_PSCI is enabled then most likely we are running on
+PSCI-enabled
+ * U-Boot which, we assume, has already taken care of configuring ARCH
+timer
+ * stuff before switching to non-secure mode.
+ */
+#if !defined(CONFIG_ARM_PSCI)

Is it required? If you see the below comment, it is already taken care by the below code.

Quoted 1 line

Is your code entering into this block, when booting the kernel in NS mode?

No, the code below seems to do the proper things. If arch timer is configured in U-Boot (enabled and programmed with correct freq) then

that code doesn't try to update it.

But "#if !defined(CONFIG_ARM_PSCI)" wraps secure_cntvoff_init() call as well, which resets CNTVOFF,

and this is something I tried to avoid the first from being executed on a CPU, which is in SVC mode (when running on top of Xen).

I was thinking whether I should wrap arch timer updating procedure as well and decided to do so, but probably I shouldn't have.

Quoted 36 lines

<snip>
/*

  87          * Update the timer if it is either not running, or is not at the
  88          * right frequency. The timer is only configurable in secure mode
  89          * so this avoids an abort if the loader started the timer and
  90          * entered the kernel in non-secure mode.
  91          */
  92
  93         if ((ioread32(base + CNTCR) & 1) == 0 ||
  94             ioread32(base + CNTFID0) != freq) {
  95                 /* Update registers with correct frequency */
  96                 iowrite32(freq, base + CNTFID0);
  97                 asm volatile("mcr p15, 0, %0, c14, c0, 0" : : "r" (freq));

98
99 /* make sure arch timer is started by setting bit 0 of CNTCR */
100 iowrite32(1, base + CNTCR);
101 }

Regards,
Biju

Quoted 13 lines
arch/arm/mach-shmobile/setup-rcar-gen2.c +1 -1
  	void __iomem *base;
  	u32 freq;
@@ -101,6 +107,7 @@ void __init rcar_gen2_timer_init(void)
  	}
 
  	iounmap(base);
+#endif /* #if !defined(CONFIG_ARM_PSCI) */
 
  	of_clk_init(NULL);
  	timer_probe();
--

2.7.4

Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#3 lore raw
In reply to Oleksandr Tyshchenko

Hi,

Quoted 51 lines · Oleksandr Tyshchenko

From: Oleksandr Tyshchenko <[email protected]>

If CONFIG_PSCI is enabled then most likely we are running on PSCI-enabled U-Boot which, we assume, has already taken care of configuring ARCH timer stuff before switching to non-secure mode.

---
A bit of context here...

We are highly interested in Renesas "Stout" board support (r8a7790) in Xen hypervisor. The reason is to have fully supported HW for performing "OSSTEST" (Xen automatic test system) on ARM32.

To reach that target we need a "generic way" for the secondary CPU cores bring up and switching them to non-secure hyp mode.
So, the PSCI as a generic well-known way to bring up CPUs, was chosen for that purpose.

You can find corresponding patches for U-Boot here:
http://u-boot.10912.n7.nabble.com/PATCH-0-3-PSCI-support-for-r8a7790-SoC-Lager-Stout-boards-td357352.html

You can find corresponding patches for Xen hypervisor here:
https://www.mail-archive.com/[email protected]/msg43332.html

To sumarize:
Together with enabling CONFIG_PSCI in shmobile_defconfig, current patch is a minimal required change needed to run mainline Linux on top of PSCI-enabled U-Boot.
There is no need to modify device tree. U-Boot will take care of inserting proper "enable-method" strings in CPU nodes.

---
  arch/arm/mach-shmobile/setup-rcar-gen2.c | 9 ++++++++-
  1 file changed, 8 insertions(+), 1 deletion(-)
arch/arm/mach-shmobile/setup-rcar-gen2.c +1 -1
index eea60b2..bac4490 100644
@@ -32,7 +32,7 @@ static const struct of_device_id cpg_matches[] __initconst = {
  	{ /* sentinel */ }
  };
 
-static unsigned int __init get_extal_freq(void)
+static unsigned int __init __maybe_unused get_extal_freq(void)
  {
  	const struct of_device_id *match;
  	struct device_node *cpg, *extal;

arch/arm/mach-shmobile/setup-rcar-gen2.c static unsigned int __init get_extal_freq(void) · line 67

Quoted 8 lines
arch/arm/mach-shmobile/setup-rcar-gen2.c +5 -0
@@ -60,6 +60,12 @@ static unsigned int __init get_extal_freq(void)
 
  void __init rcar_gen2_timer_init(void)
  {
+/*
+ * If CONFIG_PSCI is enabled then most likely we are running on PSCI-enabled
+ * U-Boot which, we assume, has already taken care of configuring ARCH timer
+ * stuff before switching to non-secure mode.
+ */

I disagree with that statement. CONFIG_ARM_PSCI may be selected in your .config for multiple reasons:
- CONFIG_XEN will selected it
- you build a .config targeting multiple platforms

Yet you may still want to boot the kernel on older U-boot. Such change would make it impossible.

If there are code that should not be called when PSCI is used, then you need to make sure the function is not called at runtime.

Cheers,

Signature · 1 line
Julien Grall
author
#4 lore raw
Quoted 1 line · Julien Grall

Hi,

Hi Julien

<snip>

Quoted 29 lines
Quoted 19 lines
arch/arm/mach-shmobile/setup-rcar-gen2.c +1 -0
index eea60b2..bac4490 100644
@@ -32,7 +32,7 @@ static const struct of_device_id cpg_matches[] 
 __initconst = {
       { /* sentinel */ }
   };
   -static unsigned int __init get_extal_freq(void)
+static unsigned int __init __maybe_unused get_extal_freq(void)
   {
       const struct of_device_id *match;

      struct device_node *cpg, *extal;

arch/arm/mach-shmobile/setup-rcar-gen2.c +2 -0
@@ -60,6 +60,12 @@ static unsigned int __init get_extal_freq(void)
     void __init rcar_gen2_timer_init(void)
   {
+/*
+ * If CONFIG_PSCI is enabled then most likely we are running on 
Quoted 5 lines

PSCI-enabled
+ * U-Boot which, we assume, has already taken care of configuring ARCH timer
+ * stuff before switching to non-secure mode.
+ */

I disagree with that statement. CONFIG_ARM_PSCI may be selected in your .config for multiple reasons:
     - CONFIG_XEN will selected it
     - you build a .config targeting multiple platforms

Quoted 3 lines

Yet you may still want to boot the kernel on older U-boot. Such change would make it impossible.

If there are code that should not be called when PSCI is used, then you need to make sure the function is not called at runtime.

I got your point.

As I answered in a previous email:

Quoted 5 lines

But "#if !defined(CONFIG_ARM_PSCI)" wraps secure_cntvoff_init() call as well, which resets CNTVOFF,

and this is something I tried to avoid the first from being executed on a CPU, which is in SVC mode (when running on top of Xen).

What would be a proper way to avoid calling secure_cntvoff_init() at runtime?

Check that we are already in non-secure mode and/or cntvoff is already 0?

Quoted 3 lines

Cheers,

Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#5 lore raw

Hi,

Quoted 39 lines · Oleksandr
Quoted 25 lines · Julien Grall
Quoted 13 lines
arch/arm/mach-shmobile/setup-rcar-gen2.c +1 -0
index eea60b2..bac4490 100644
@@ -32,7 +32,7 @@ static const struct of_device_id cpg_matches[] __initconst = {
       { /* sentinel */ }
   };
   -static unsigned int __init get_extal_freq(void)
+static unsigned int __init __maybe_unused get_extal_freq(void)
   {
       const struct of_device_id *match;
       struct device_node *cpg, *extal;

arch/arm/mach-shmobile/setup-rcar-gen2.c static unsigned int __init get_extal_freq(void) · line 66

Quoted 7 lines
arch/arm/mach-shmobile/setup-rcar-gen2.c +5 -0
@@ -60,6 +60,12 @@ static unsigned int __init get_extal_freq(void)
     void __init rcar_gen2_timer_init(void)
   {
+/*
+ * If CONFIG_PSCI is enabled then most likely we are running on PSCI-enabled
+ * U-Boot which, we assume, has already taken care of configuring ARCH timer
+ * stuff before switching to non-secure mode.
+ */

I disagree with that statement. CONFIG_ARM_PSCI may be selected in your .config for multiple reasons:
     - CONFIG_XEN will selected it
     - you build a .config targeting multiple platforms

Quoted 3 lines

Yet you may still want to boot the kernel on older U-boot. Such change would make it impossible.

If there are code that should not be called when PSCI is used, then you need to make sure the function is not called at runtime.

I got your point.

As I answered in a previous email:

Quoted 2 lines

But "#if !defined(CONFIG_ARM_PSCI)" wraps secure_cntvoff_init() call as well, which resets CNTVOFF,

Quoted 6 lines
Quoted 2 lines

and this is something I tried to avoid the first from being executed on a CPU, which is in SVC mode (when running on top of Xen).

What would be a proper way to avoid calling secure_cntvoff_init() at runtime?

Check that we are already in non-secure mode and/or cntvoff is already 0?

CNTVOFF is controlled by the hypervisor, so you cannot access it from Dom0.

If none of this code should be necessary when PSCI is enabled, then you can check the presence of PSCI (see psci_smp_available()).

Another solution is to check whether we are running on xen (see xen_domain()).

In any case, I would still surround the two checks because the "armgcnt" should really not be touched by Dom0.

Cheers,

Signature · 1 line
Julien Grall
author
#6 lore raw
Quoted 1 line · Julien Grall

Hi,

Hi Julien

Quoted 10 lines
Quoted 10 lines

As I answered in a previous email:

 > But "#if !defined(CONFIG_ARM_PSCI)" wraps secure_cntvoff_init() call as > well, which resets CNTVOFF,

 > and this is something I tried to avoid the first from being executed on > a CPU, which is in SVC mode (when running on top of Xen).

Quoted 7 lines
Quoted 3 lines

What would be a proper way to avoid calling secure_cntvoff_init() at runtime?

Check that we are already in non-secure mode and/or cntvoff is already 0?

CNTVOFF is controlled by the hypervisor, so you cannot access it from Dom0.

If none of this code should be necessary when PSCI is enabled, then you can check the presence of PSCI (see psci_smp_available()).

If I am not mistaken, U-Boot resets CNTVOFF before leaving monitor mode (switching to non-secure state). This happens if "support for booting in non-secure mode" is enabled.

This is a mandatory option for PSCI support as well as for support for hardware virtualization.

So the presence of PSCI would be an indicator, if CNTVOFF was configured in U-Boot and we don't need to.

From other hand, Gen2 Linux may be booted in non-secure mode, without PSCI support...

I was thinking about checking the presence of "enable-method = psci" in CPU node, but probably psci_smp_available() is better option.

Also, there is a specific device-tree property "arm,cpu-registers-not-fw-configured" [1], but I am not sure whether it is appropriate to use here.

Quoted 4 lines

Another solution is to check whether we are running on xen (see xen_domain()).

In any case, I would still surround the two checks because the "armgcnt" should really not be touched by Dom0.

Agree.

I will add two checks in V2.

Thank you!

[1]

https://elixir.bootlin.com/linux/v5.1-rc6/source/Documentation/devicetree/bindings/timer/arm,arch_timer.yaml#L68

Quoted 3 lines

Cheers,

Signature · 3 lines
Regards,

Oleksandr Tyshchenko