[PATCH RFC 1/4] arm64: kernel: implement DT based idle states infrastructure
Lorenzo Pieralisi
lorenzo.pieralisi at arm.com
Tue Mar 18 14:08:57 EDT 2014
Hi Rob,
thanks for reviewing.
On Tue, Mar 18, 2014 at 01:27:29PM +0000, Rob Herring wrote:
> On Tue, Mar 18, 2014 at 5:20 AM, Lorenzo Pieralisi
> <lorenzo.pieralisi at arm.com> wrote:
> > On most common ARM systems, the low-power states a CPU can be put into are
> > not discoverable in HW and require device tree bindings to describe
> > the respective power domains, power down protocol and idle states parameters.
> >
> > In order to enable DT based idle states and configure idle drivers, this
> > patch implements the bulk infrastructure required to parse the device tree
> > idle states bindings and functions to initizialize idle driver and protocol
> > back-ends.
> >
> > Protocol back-ends (eg PSCI) must register a protocol initializer with
> > the idle state parser so that upon protocol detection, the parsing code
> > can call the back-end infrastructure to complete the idle driver
> > initialization.
> >
> > Idle state index 0 is always initialized, ie always considered present
> > on all ARM platforms.
> >
> > Code that initializes idle states checks the CPU idle driver cpumask so
> > that multiple CPU idle drivers can be initialized through it in the
> > kernel. The CPU idle driver cpumask defines which idle states should be
> > considered valid for the driver, ie idle states that are valid on a set
> > of cpus the idle driver manages.
> >
> > Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi at arm.com>
> > ---
> > arch/arm64/Kconfig | 4 +
> > arch/arm64/include/asm/idle_states.h | 20 ++
> > arch/arm64/kernel/Makefile | 1 +
> > arch/arm64/kernel/idle_states.c | 397 +++++++++++++++++++++++++++++++++++
>
> This all belongs in drivers/cpuidle either as part of your driver or
> as library calls the driver can use as it is very obvious it is
> dependent on cpuidle.
Yes, and it is something I thought about before posting. There are a
couple of dependencies to be managed though (PSCI and cpu_suspend API),
I can probably make the build dependent on ARM64 for the moment.
I will look into this.
> > 4 files changed, 422 insertions(+)
> > create mode 100644 arch/arm64/include/asm/idle_states.h
> > create mode 100644 arch/arm64/kernel/idle_states.c
> >
> > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > index 27bbcfc..3132572 100644
> > --- a/arch/arm64/Kconfig
> > +++ b/arch/arm64/Kconfig
> > @@ -295,6 +295,10 @@ config ARCH_SUSPEND_POSSIBLE
> > config ARM64_CPU_SUSPEND
> > def_bool PM_SLEEP
> >
> > +config ARM64_IDLE_STATES
>
> Idle states apply to ARM as well, right?
Eh, yes with a couple of twists, see above.
> OF_IDLE_STATES?
Seems sensible.
> > + def_bool CPU_IDLE
> > + select ARM64_CPU_SUSPEND
>
> This should probably get renamed to ARCH_CPU_SUSPEND. There's no need
> for different names for arm and arm64.
Yes I think that makes sense, I will move parsing code to drivers/cpuidle and
check if the CONFIG rename might cause issues.
[...]
> > +static struct idle_state idle_states[CPUIDLE_STATE_MAX] __initdata;
> > +typedef int (*idle_states_initializer)(struct cpumask *, struct idle_state *,
> > + unsigned int);
> > +
> > +struct protocol_init {
> > + const char *prot;
> > + idle_states_initializer prot_init;
> > +};
> > +
> > +static const struct protocol_init protocols[] __initconst = {
> > + {}
> > +};
> > +
> > +static __init const struct protocol_init *get_protocol(const char *str)
> > +{
>
> This needs a better name or to be removed until you actually need it.
> What protocols do you expect to have? This kind of looks like GUFI
> trying to do DT vs. ACPI backends.
Better name probably, but I do now want to hardcode PSCI either (patch 2 adds
the PSCI backend initializer), I am not sure that would make people extremely
happy and this code can be reused by different back-end implementations with
nary a problem.
Code has to be generic enough to call initialization for different suspend
backends, I do not think I am asking too much here.
> > + int i;
> > +
> > + if (!str)
> > + return NULL;
> > +
> > + for (i = 0; protocols[i].prot; i++)
> > + if (!strcmp(protocols[i].prot, str))
> > + return &protocols[i];
> > +
> > + return NULL;
> > +}
> > +
> > +static void __init idle_state_cpu_mask(int cpu, struct idle_state *idle_state,
> > + struct device_node *cn)
> > +{
> > + int i = 0;
> > + struct device_node *cpu_state;
> > +
> > + do {
> > + cpu_state = of_parse_phandle(cn, "cpu-idle-states", i++);
> > + if (cpu_state && idle_state->node == cpu_state)
> > + cpumask_set_cpu(cpu, &idle_state->cpus);
> > + of_node_put(cpu_state);
> > + } while (cpu_state);
> > +}
> > +
> > +static int __init parse_idle_states_node(struct device_node *parent, int cnt,
> > + const cpumask_t *cpus)
> > +{
> > + struct device_node *state;
>
> node or np or state_node so it is clear this is a device_node.
Ok.
> > + struct idle_state *idle_state;
> > + int cpu;
> > +
> > + for_each_child_of_node(parent, state) {
> > +
> > + if (!of_device_is_compatible(state, "arm,idle-state")) {
> > + pr_warn(" %s has child nodes that are not idle states\n",
> > + parent->full_name);
> > + continue;
> > + }
> > +
> > + idle_state = &idle_states[cnt];
> > +
> > + pr_debug(" * %s...\n", state->full_name);
> > +
> > + idle_state->node = state;
> > + /*
> > + * Check cpus on which the idle state is valid
> > + */
> > + for_each_possible_cpu(cpu) {
> > + struct device_node *cn;
> > +
> > + cn = of_get_cpu_node(cpu, NULL);
> > + if (!cn) {
> > + pr_err("missing device node for CPU %d\n", cpu);
> > + continue;
> > + }
> > + idle_state_cpu_mask(cpu, idle_state, cn);
> > + }
> > +
> > + /*
> > + * The driver cpumask is not a subset of cpus on which the
> > + * idle state is valid, hence the idle state is skipped for
> > + * this driver.
> > + */
> > + if (!cpumask_subset(cpus, &idle_state->cpus))
> > + continue;
> > +
> > + if (of_property_read_u32(state, "index", &idle_state->index)) {
> > + pr_debug(" * %s missing index property\n",
> > + state->full_name);
> > + continue;
> > + }
> > +
> > + if (of_property_read_u32(state, "entry-latency-us",
> > + &idle_state->entry_latency)) {
> > + pr_debug(" * %s missing entry latency property\n",
> > + state->full_name);
> > + continue;
> > + }
> > +
> > + if (of_property_read_u32(state, "exit-latency-us",
> > + &idle_state->exit_latency)) {
> > + pr_debug(" * %s missing exit latency property\n",
> > + state->full_name);
> > + continue;
> > + }
> > +
> > + if (of_property_read_u32(state, "min-residency-us",
> > + &idle_state->min_residency)) {
> > + pr_debug(" * %s missing min-residency property\n",
> > + state->full_name);
> > + continue;
> > + }
> > +
> > + if (of_property_read_u32(state, "entry-method-param",
> > + &idle_state->param)) {
> > + pr_debug(" * %s missing entry-method-param property\n",
> > + state->full_name);
> > + continue;
> > + }
> > +
> > + if (++cnt == CPUIDLE_STATE_MAX) {
> > + pr_warn("Number of parsed states equal static CPU idle state limit\n");
> > + of_node_put(state);
> > + break;
> > + }
> > + }
> > +
> > + return cnt;
> > +}
> > +
> > +static int __init parse_idle_states(struct device_node *root, int cnt,
> > + const cpumask_t *cpus)
> > +{
> > + int head_idx, curr_idx;
> > + struct device_node *curr = root;
> > +
> > + /*
> > + * Breadth-first DT idle states parsing
> > + *
> > + * Sweep idle states level in the device tree and use the
> > + * idle_states array to stash the visited nodes, as a queue.
> > + *
> > + * parse_idle_states_node() updates the idle_states array by
> > + * initializing entries, stashing the device tree node for the
> > + * corresponding state (struct idle_state.node) and incrementing
> > + * the idle states counter that is returned so that curr_idx is
> > + * kept up-to-date while descending into tree levels.
> > + *
> > + * Store the initial counter head_idx and curr_idx and use head_idx
> > + * as a queue of node indices to be visited.
> > + *
> > + * When we reach the max number of CPU idle states or
> > + * head_idx == curr_idx (empty nodes queue) we are done.
> > + */
> > + head_idx = curr_idx = cnt;
> > +
> > + do {
> > + curr_idx = parse_idle_states_node(curr, curr_idx, cpus);
> > + if (curr_idx == CPUIDLE_STATE_MAX || head_idx == curr_idx)
> > + break;
> > + /*
> > + * idle_states array is updated by parse_idle_states_node(),
> > + * we can use the initialized states as a queue of nodes
> > + * that need to be checked for their idle states siblings.
> > + * head_idx works as a pointer into the queue to get the
> > + * next node to be parsed.
> > + */
> > + curr = idle_states[head_idx++].node;
> > + } while (curr);
>
> I still object to index property and this is why. You need to be able
> to determine state order by actual h/w properties. That is what you
> are doing in your head when you define the indexes.
>
> You really want a linked list here that you can sort as you go and not
> care what order you parse DT nodes. Not to mention you don't know how
> many states you will have.
This code does not care about the order of nodes, the index is just there
to keep track of nodes that have still to be parsed. Sorting is done later,
using the index property (totally unrelated to the {head/curr}_idx) which I
understand is frowned upon in DT world (but in this case I think it could be
accepted, certainly it would make my life easier).
Having said that, I like the idea of implementing it with a linked list and
sorting states while parsing them. I will remove that index property and
replace it with an actual hw property: power_consumption ? Or should I just
use min_residency (the higher the required residency the deeper the idle
state) ? Defining the power consumption (or better savings) for a state is
an _outright_ can of worms, that's why using an index is easier.
Thoughts ?
Thanks,
Lorenzo
More information about the linux-arm-kernel
mailing list