diff mbox

[v5,4/8] arm64: add PSCI CPU_SUSPEND based cpu_suspend support

Message ID 1403705421-17597-5-git-send-email-lorenzo.pieralisi@arm.com (mailing list archive)
State New, archived
Headers show

Commit Message

Lorenzo Pieralisi June 25, 2014, 2:10 p.m. UTC
This patch implements the cpu_suspend cpu operations method through
the PSCI CPU_SUSPEND API. The PSCI implementation translates the idle state
index passed by the cpu_suspend core call into a valid PSCI state according to
the PSCI states initialized at boot by the PSCI suspend backend.

Entry point is set to cpu_resume physical address, that represents the
default kernel execution address following a CPU reset.

Idle state indices missing a DT node description are initialized to power
state standby WFI so that if called by the idle driver they provide the
default behaviour.

Reviewed-by: Sebastian Capella <sebcape@gmail.com>
Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
---
 arch/arm64/include/asm/psci.h |   4 ++
 arch/arm64/kernel/psci.c      | 103 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 107 insertions(+)

Comments

Mark Rutland June 25, 2014, 4:09 p.m. UTC | #1
On Wed, Jun 25, 2014 at 03:10:17PM +0100, Lorenzo Pieralisi wrote:
> This patch implements the cpu_suspend cpu operations method through
> the PSCI CPU_SUSPEND API. The PSCI implementation translates the idle state
> index passed by the cpu_suspend core call into a valid PSCI state according to
> the PSCI states initialized at boot by the PSCI suspend backend.
> 
> Entry point is set to cpu_resume physical address, that represents the
> default kernel execution address following a CPU reset.
> 
> Idle state indices missing a DT node description are initialized to power
> state standby WFI so that if called by the idle driver they provide the
> default behaviour.
> 
> Reviewed-by: Sebastian Capella <sebcape@gmail.com>
> Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
> ---
>  arch/arm64/include/asm/psci.h |   4 ++
>  arch/arm64/kernel/psci.c      | 103 ++++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 107 insertions(+)

[...]

> +static void psci_power_state_unpack(u32 power_state,
> +				    struct psci_power_state *state)
> +{
> +	state->id = (power_state & PSCI_0_2_POWER_STATE_ID_MASK) >>
> +			PSCI_0_2_POWER_STATE_ID_SHIFT;
> +	state->type = (power_state & PSCI_0_2_POWER_STATE_TYPE_MASK) >>
> +			PSCI_0_2_POWER_STATE_TYPE_SHIFT;
> +	state->affinity_level =
> +			(power_state & PSCI_0_2_POWER_STATE_AFFL_MASK) >>
> +			PSCI_0_2_POWER_STATE_AFFL_SHIFT;
> +}

Is this valid for PSCI versions prior to 0.2?

>  /*
>   * The following two functions are invoked via the invoke_psci_fn pointer
>   * and will not be inlined, allowing us to piggyback on the AAPCS.
> @@ -199,6 +216,77 @@ static int psci_migrate_info_type(void)
>  	return err;
>  }
>  
> +int __init psci_dt_register_idle_states(struct cpuidle_driver *drv,
> +					struct device_node *state_nodes[])
> +{
> +	int cpu, i;

Perhaps unsigned int? You print i with %u below.

> +	for (i = 0; i < drv->state_count; i++) {
> +		u32 psci_power_state;
> +
> +		if (!state_nodes[i]) {
> +			/*
> +			 * An index with a missing node pointer falls back to
> +			 * simple STANDBYWFI
> +			 */
> +			psci_states[i].type = PSCI_POWER_STATE_TYPE_STANDBY;
> +			continue;
> +		}

Does this make sense? Are there any limitations on which state nodes
could be missing?

> +
> +		if (of_property_read_u32(state_nodes[i], "entry-method-param",
> +					 &psci_power_state)) {
> +			pr_warn(" * %s missing entry-method-param property\n",
> +				state_nodes[i]->full_name);
> +			/*
> +			 * If entry-method-param property is missing, fall
> +			 * back to STANDBYWFI state
> +			 */
> +			psci_states[i].type = PSCI_POWER_STATE_TYPE_STANDBY;
> +			continue;

Surely we want to throw away these states instead?

Otherwise we can get into a mess like:

psci_states[0] => low power state
psci_states[1] => lower power state
psci_states[2] => WFI / not low power
psci_states[3] => lowest power state

Where power usage and latency would jump around rather than follow
monotonic patterns.

Thanks,
Mark.
Geoff Levand June 25, 2014, 8:52 p.m. UTC | #2
Hi Lorenzo,

On Wed, 2014-06-25 at 15:10 +0100, Lorenzo Pieralisi wrote:
> diff --git a/arch/arm64/kernel/psci.c b/arch/arm64/kernel/psci.c

> +	/*
> +	 * This is belt-and-braces: make sure that if the idle
> +	 * specified protocol is psci, the cpu_ops have been
> +	 * initialized to psci operations. Anything else is
> +	 * a recipe for mayhem.
> +	 */
> +	for_each_cpu(cpu, drv->cpumask) {
> +		cpu_ops_ptr = cpu_ops[cpu];
> +		if (WARN_ON(!cpu_ops_ptr || strcmp(cpu_ops_ptr->name, "psci")))
> +			return -EOPNOTSUPP;
> +	}

I'm not sure how drv->cpumask is setup, but if a system has mixed enable
methods, say some cpus 'spin-table' and some 'psci', will this give a
false error?

If drv->cpumask should only include 'psci' cpus, then should this be a
BUG()?

> +
> +	psci_states = kcalloc(drv->state_count, sizeof(*psci_states),
> +			      GFP_KERNEL);
> +
> +	if (!psci_states) {
> +		pr_warn("psci idle state allocation failed\n");
> +		return -ENOMEM;
> +	}
> +
> +	for_each_cpu(cpu, drv->cpumask) {
> +		if (per_cpu(psci_power_state, cpu)) {
> +			pr_warn("idle states already initialized on cpu %u\n",
> +				cpu);

This seems like an implementation problem, if so, shouldn't this be
pr_debug()?


>  #endif
>  
> +#ifdef CONFIG_ARM64_CPU_SUSPEND
> +static int cpu_psci_cpu_suspend(unsigned long index)
> +{
> +	struct psci_power_state *state = __get_cpu_var(psci_power_state);
> +
> +	if (!state)
> +		return -EOPNOTSUPP;
> +
> +	return psci_ops.cpu_suspend(state[index], virt_to_phys(cpu_resume));
> +}
> +#endif

Why not put a __maybe_unused attribute on cpu_psci_cpu_suspend() and
remove the preprocessor conditional.  That way this code will always be
compiled, and with therefor always get a build test.  The linker should
strip out the unused code when CONFIG_ARM64_CPU_SUSPEND=n and the code
below is not compiled. 

> +#ifdef CONFIG_ARM64_CPU_SUSPEND
> +	.cpu_suspend	= cpu_psci_cpu_suspend,
> +#endif
>  };

-Geoff
Lorenzo Pieralisi June 26, 2014, 11:23 a.m. UTC | #3
On Wed, Jun 25, 2014 at 05:09:11PM +0100, Mark Rutland wrote:
> On Wed, Jun 25, 2014 at 03:10:17PM +0100, Lorenzo Pieralisi wrote:
> > This patch implements the cpu_suspend cpu operations method through
> > the PSCI CPU_SUSPEND API. The PSCI implementation translates the idle state
> > index passed by the cpu_suspend core call into a valid PSCI state according to
> > the PSCI states initialized at boot by the PSCI suspend backend.
> > 
> > Entry point is set to cpu_resume physical address, that represents the
> > default kernel execution address following a CPU reset.
> > 
> > Idle state indices missing a DT node description are initialized to power
> > state standby WFI so that if called by the idle driver they provide the
> > default behaviour.
> > 
> > Reviewed-by: Sebastian Capella <sebcape@gmail.com>
> > Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
> > ---
> >  arch/arm64/include/asm/psci.h |   4 ++
> >  arch/arm64/kernel/psci.c      | 103 ++++++++++++++++++++++++++++++++++++++++++
> >  2 files changed, 107 insertions(+)
> 
> [...]
> 
> > +static void psci_power_state_unpack(u32 power_state,
> > +				    struct psci_power_state *state)
> > +{
> > +	state->id = (power_state & PSCI_0_2_POWER_STATE_ID_MASK) >>
> > +			PSCI_0_2_POWER_STATE_ID_SHIFT;
> > +	state->type = (power_state & PSCI_0_2_POWER_STATE_TYPE_MASK) >>
> > +			PSCI_0_2_POWER_STATE_TYPE_SHIFT;
> > +	state->affinity_level =
> > +			(power_state & PSCI_0_2_POWER_STATE_AFFL_MASK) >>
> > +			PSCI_0_2_POWER_STATE_AFFL_SHIFT;
> > +}
> 
> Is this valid for PSCI versions prior to 0.2?

Yes, it should, as for the packing function.

> >  /*
> >   * The following two functions are invoked via the invoke_psci_fn pointer
> >   * and will not be inlined, allowing us to piggyback on the AAPCS.
> > @@ -199,6 +216,77 @@ static int psci_migrate_info_type(void)
> >  	return err;
> >  }
> >  
> > +int __init psci_dt_register_idle_states(struct cpuidle_driver *drv,
> > +					struct device_node *state_nodes[])
> > +{
> > +	int cpu, i;
> 
> Perhaps unsigned int? You print i with %u below.

Yes.

> > +	for (i = 0; i < drv->state_count; i++) {
> > +		u32 psci_power_state;
> > +
> > +		if (!state_nodes[i]) {
> > +			/*
> > +			 * An index with a missing node pointer falls back to
> > +			 * simple STANDBYWFI
> > +			 */
> > +			psci_states[i].type = PSCI_POWER_STATE_TYPE_STANDBY;
> > +			continue;
> > +		}
> 
> Does this make sense? Are there any limitations on which state nodes
> could be missing?

I think the check is overkill, you are right.

> > +
> > +		if (of_property_read_u32(state_nodes[i], "entry-method-param",
> > +					 &psci_power_state)) {
> > +			pr_warn(" * %s missing entry-method-param property\n",
> > +				state_nodes[i]->full_name);
> > +			/*
> > +			 * If entry-method-param property is missing, fall
> > +			 * back to STANDBYWFI state
> > +			 */
> > +			psci_states[i].type = PSCI_POWER_STATE_TYPE_STANDBY;
> > +			continue;
> 
> Surely we want to throw away these states instead?
> 
> Otherwise we can get into a mess like:
> 
> psci_states[0] => low power state
> psci_states[1] => lower power state
> psci_states[2] => WFI / not low power
> psci_states[3] => lowest power state
> 
> Where power usage and latency would jump around rather than follow
> monotonic patterns.

I do not think that's a problem by itself, but honestly I think you have
a point. It is better to barf, throw away the states and avoid initializing
CPUidle to force a firmware update than keep going with a state that is
actually not doing what it probably was designed for, I just tried to be too
accommodating on this.

Thanks,
Lorenzo
Lorenzo Pieralisi June 26, 2014, 4:55 p.m. UTC | #4
Hi Geoff,

On Wed, Jun 25, 2014 at 09:52:00PM +0100, Geoff Levand wrote:
> Hi Lorenzo,
> 
> On Wed, 2014-06-25 at 15:10 +0100, Lorenzo Pieralisi wrote:
> > diff --git a/arch/arm64/kernel/psci.c b/arch/arm64/kernel/psci.c
> 
> > +	/*
> > +	 * This is belt-and-braces: make sure that if the idle
> > +	 * specified protocol is psci, the cpu_ops have been
> > +	 * initialized to psci operations. Anything else is
> > +	 * a recipe for mayhem.
> > +	 */
> > +	for_each_cpu(cpu, drv->cpumask) {
> > +		cpu_ops_ptr = cpu_ops[cpu];
> > +		if (WARN_ON(!cpu_ops_ptr || strcmp(cpu_ops_ptr->name, "psci")))
> > +			return -EOPNOTSUPP;
> > +	}
> 
> I'm not sure how drv->cpumask is setup, but if a system has mixed enable
> methods, say some cpus 'spin-table' and some 'psci', will this give a
> false error?

I do not think that's a false error. If the idle states specify an
entry-method == psci, and cpu_ops for some cpus are not set to PSCI,
obviously because the enable-method specified that, that's a firmware bug.

> If drv->cpumask should only include 'psci' cpus, then should this be a
> BUG()?

Ok, if we got here, it is because the idle-states entry-method was set
to PSCI. Now, if any of the CPUs in the driver mask has a cpu_ops
pointer != PSCI, we have a problem and we should warn on that. I do
not think that justifies a BUG_ON, but that's one of those things, it is
debatable.

Question is whether the check should also be carried out at cpu_ops
initialization (ie to check for mixed cpu_ops), for certain if the
idle states entry-method is PSCI and cpu_ops != PSCI we should
WARN/BUG on that. Or embed this idle state parameters initialization at
cpu_ops init (see other thread you started) so that we can kill two
birds with one stone.

> > +
> > +	psci_states = kcalloc(drv->state_count, sizeof(*psci_states),
> > +			      GFP_KERNEL);
> > +
> > +	if (!psci_states) {
> > +		pr_warn("psci idle state allocation failed\n");
> > +		return -ENOMEM;
> > +	}
> > +
> > +	for_each_cpu(cpu, drv->cpumask) {
> > +		if (per_cpu(psci_power_state, cpu)) {
> > +			pr_warn("idle states already initialized on cpu %u\n",
> > +				cpu);
> 
> This seems like an implementation problem, if so, shouldn't this be
> pr_debug()?

Maybe, I will give it some thought.

> >  #endif
> >  
> > +#ifdef CONFIG_ARM64_CPU_SUSPEND
> > +static int cpu_psci_cpu_suspend(unsigned long index)
> > +{
> > +	struct psci_power_state *state = __get_cpu_var(psci_power_state);
> > +
> > +	if (!state)
> > +		return -EOPNOTSUPP;
> > +
> > +	return psci_ops.cpu_suspend(state[index], virt_to_phys(cpu_resume));
> > +}
> > +#endif
> 
> Why not put a __maybe_unused attribute on cpu_psci_cpu_suspend() and
> remove the preprocessor conditional.  That way this code will always be
> compiled, and with therefor always get a build test.  The linker should
> strip out the unused code when CONFIG_ARM64_CPU_SUSPEND=n and the code
> below is not compiled. 

It can make sense, yes.

Thanks,
Lorenzo
diff mbox

Patch

diff --git a/arch/arm64/include/asm/psci.h b/arch/arm64/include/asm/psci.h
index e5312ea..16c1351 100644
--- a/arch/arm64/include/asm/psci.h
+++ b/arch/arm64/include/asm/psci.h
@@ -14,6 +14,10 @@ 
 #ifndef __ASM_PSCI_H
 #define __ASM_PSCI_H
 
+struct cpuidle_driver;
 int psci_init(void);
 
+int __init psci_dt_register_idle_states(struct cpuidle_driver *,
+					struct device_node *[]);
+
 #endif /* __ASM_PSCI_H */
diff --git a/arch/arm64/kernel/psci.c b/arch/arm64/kernel/psci.c
index 9e9798f..f708bcc 100644
--- a/arch/arm64/kernel/psci.c
+++ b/arch/arm64/kernel/psci.c
@@ -15,12 +15,14 @@ 
 
 #define pr_fmt(fmt) "psci: " fmt
 
+#include <linux/cpuidle.h>
 #include <linux/init.h>
 #include <linux/of.h>
 #include <linux/smp.h>
 #include <linux/reboot.h>
 #include <linux/pm.h>
 #include <linux/delay.h>
+#include <linux/slab.h>
 #include <uapi/linux/psci.h>
 
 #include <asm/compiler.h>
@@ -28,6 +30,7 @@ 
 #include <asm/errno.h>
 #include <asm/psci.h>
 #include <asm/smp_plat.h>
+#include <asm/suspend.h>
 #include <asm/system_misc.h>
 
 #define PSCI_POWER_STATE_TYPE_STANDBY		0
@@ -65,6 +68,8 @@  enum psci_function {
 	PSCI_FN_MAX,
 };
 
+static DEFINE_PER_CPU_READ_MOSTLY(struct psci_power_state *, psci_power_state);
+
 static u32 psci_function_id[PSCI_FN_MAX];
 
 static int psci_to_linux_errno(int errno)
@@ -93,6 +98,18 @@  static u32 psci_power_state_pack(struct psci_power_state state)
 		 & PSCI_0_2_POWER_STATE_AFFL_MASK);
 }
 
+static void psci_power_state_unpack(u32 power_state,
+				    struct psci_power_state *state)
+{
+	state->id = (power_state & PSCI_0_2_POWER_STATE_ID_MASK) >>
+			PSCI_0_2_POWER_STATE_ID_SHIFT;
+	state->type = (power_state & PSCI_0_2_POWER_STATE_TYPE_MASK) >>
+			PSCI_0_2_POWER_STATE_TYPE_SHIFT;
+	state->affinity_level =
+			(power_state & PSCI_0_2_POWER_STATE_AFFL_MASK) >>
+			PSCI_0_2_POWER_STATE_AFFL_SHIFT;
+}
+
 /*
  * The following two functions are invoked via the invoke_psci_fn pointer
  * and will not be inlined, allowing us to piggyback on the AAPCS.
@@ -199,6 +216,77 @@  static int psci_migrate_info_type(void)
 	return err;
 }
 
+int __init psci_dt_register_idle_states(struct cpuidle_driver *drv,
+					struct device_node *state_nodes[])
+{
+	int cpu, i;
+	struct psci_power_state *psci_states;
+	const struct cpu_operations *cpu_ops_ptr;
+
+	if (!state_nodes)
+		return -EINVAL;
+	/*
+	 * This is belt-and-braces: make sure that if the idle
+	 * specified protocol is psci, the cpu_ops have been
+	 * initialized to psci operations. Anything else is
+	 * a recipe for mayhem.
+	 */
+	for_each_cpu(cpu, drv->cpumask) {
+		cpu_ops_ptr = cpu_ops[cpu];
+		if (WARN_ON(!cpu_ops_ptr || strcmp(cpu_ops_ptr->name, "psci")))
+			return -EOPNOTSUPP;
+	}
+
+	psci_states = kcalloc(drv->state_count, sizeof(*psci_states),
+			      GFP_KERNEL);
+
+	if (!psci_states) {
+		pr_warn("psci idle state allocation failed\n");
+		return -ENOMEM;
+	}
+
+	for_each_cpu(cpu, drv->cpumask) {
+		if (per_cpu(psci_power_state, cpu)) {
+			pr_warn("idle states already initialized on cpu %u\n",
+				cpu);
+			continue;
+		}
+		per_cpu(psci_power_state, cpu) = psci_states;
+	}
+
+
+	for (i = 0; i < drv->state_count; i++) {
+		u32 psci_power_state;
+
+		if (!state_nodes[i]) {
+			/*
+			 * An index with a missing node pointer falls back to
+			 * simple STANDBYWFI
+			 */
+			psci_states[i].type = PSCI_POWER_STATE_TYPE_STANDBY;
+			continue;
+		}
+
+		if (of_property_read_u32(state_nodes[i], "entry-method-param",
+					 &psci_power_state)) {
+			pr_warn(" * %s missing entry-method-param property\n",
+				state_nodes[i]->full_name);
+			/*
+			 * If entry-method-param property is missing, fall
+			 * back to STANDBYWFI state
+			 */
+			psci_states[i].type = PSCI_POWER_STATE_TYPE_STANDBY;
+			continue;
+		}
+
+		pr_debug("psci-power-state %#x index %u\n", psci_power_state,
+							    i);
+		psci_power_state_unpack(psci_power_state, &psci_states[i]);
+	}
+
+	return 0;
+}
+
 static int get_set_conduit_method(struct device_node *np)
 {
 	const char *method;
@@ -435,6 +523,18 @@  static int cpu_psci_cpu_kill(unsigned int cpu)
 }
 #endif
 
+#ifdef CONFIG_ARM64_CPU_SUSPEND
+static int cpu_psci_cpu_suspend(unsigned long index)
+{
+	struct psci_power_state *state = __get_cpu_var(psci_power_state);
+
+	if (!state)
+		return -EOPNOTSUPP;
+
+	return psci_ops.cpu_suspend(state[index], virt_to_phys(cpu_resume));
+}
+#endif
+
 const struct cpu_operations cpu_psci_ops = {
 	.name		= "psci",
 	.cpu_init	= cpu_psci_cpu_init,
@@ -445,6 +545,9 @@  const struct cpu_operations cpu_psci_ops = {
 	.cpu_die	= cpu_psci_cpu_die,
 	.cpu_kill	= cpu_psci_cpu_kill,
 #endif
+#ifdef CONFIG_ARM64_CPU_SUSPEND
+	.cpu_suspend	= cpu_psci_cpu_suspend,
+#endif
 };
 
 #endif