2019-05-27 14:55:06 +08:00
|
|
|
// SPDX-License-Identifier: GPL-2.0-or-later
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
/*
|
|
|
|
* coupled.c - helper functions to enter the same idle state on multiple cpus
|
|
|
|
*
|
|
|
|
* Copyright (c) 2011 Google, Inc.
|
|
|
|
*
|
|
|
|
* Author: Colin Cross <ccross@android.com>
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/kernel.h>
|
|
|
|
#include <linux/cpu.h>
|
|
|
|
#include <linux/cpuidle.h>
|
|
|
|
#include <linux/mutex.h>
|
|
|
|
#include <linux/sched.h>
|
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/spinlock.h>
|
|
|
|
|
|
|
|
#include "cpuidle.h"
|
|
|
|
|
|
|
|
/**
|
|
|
|
* DOC: Coupled cpuidle states
|
|
|
|
*
|
|
|
|
* On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
|
|
|
|
* cpus cannot be independently powered down, either due to
|
|
|
|
* sequencing restrictions (on Tegra 2, cpu 0 must be the last to
|
|
|
|
* power down), or due to HW bugs (on OMAP4460, a cpu powering up
|
|
|
|
* will corrupt the gic state unless the other cpu runs a work
|
|
|
|
* around). Each cpu has a power state that it can enter without
|
|
|
|
* coordinating with the other cpu (usually Wait For Interrupt, or
|
|
|
|
* WFI), and one or more "coupled" power states that affect blocks
|
|
|
|
* shared between the cpus (L2 cache, interrupt controller, and
|
|
|
|
* sometimes the whole SoC). Entering a coupled power state must
|
|
|
|
* be tightly controlled on both cpus.
|
|
|
|
*
|
|
|
|
* This file implements a solution, where each cpu will wait in the
|
|
|
|
* WFI state until all cpus are ready to enter a coupled state, at
|
|
|
|
* which point the coupled state function will be called on all
|
|
|
|
* cpus at approximately the same time.
|
|
|
|
*
|
|
|
|
* Once all cpus are ready to enter idle, they are woken by an smp
|
|
|
|
* cross call. At this point, there is a chance that one of the
|
|
|
|
* cpus will find work to do, and choose not to enter idle. A
|
|
|
|
* final pass is needed to guarantee that all cpus will call the
|
|
|
|
* power state enter function at the same time. During this pass,
|
|
|
|
* each cpu will increment the ready counter, and continue once the
|
|
|
|
* ready counter matches the number of online coupled cpus. If any
|
|
|
|
* cpu exits idle, the other cpus will decrement their counter and
|
|
|
|
* retry.
|
|
|
|
*
|
|
|
|
* requested_state stores the deepest coupled idle state each cpu
|
|
|
|
* is ready for. It is assumed that the states are indexed from
|
|
|
|
* shallowest (highest power, lowest exit latency) to deepest
|
|
|
|
* (lowest power, highest exit latency). The requested_state
|
|
|
|
* variable is not locked. It is only written from the cpu that
|
|
|
|
* it stores (or by the on/offlining cpu if that cpu is offline),
|
|
|
|
* and only read after all the cpus are ready for the coupled idle
|
|
|
|
* state are are no longer updating it.
|
|
|
|
*
|
|
|
|
* Three atomic counters are used. alive_count tracks the number
|
|
|
|
* of cpus in the coupled set that are currently or soon will be
|
|
|
|
* online. waiting_count tracks the number of cpus that are in
|
|
|
|
* the waiting loop, in the ready loop, or in the coupled idle state.
|
|
|
|
* ready_count tracks the number of cpus that are in the ready loop
|
|
|
|
* or in the coupled idle state.
|
|
|
|
*
|
|
|
|
* To use coupled cpuidle states, a cpuidle driver must:
|
|
|
|
*
|
|
|
|
* Set struct cpuidle_device.coupled_cpus to the mask of all
|
|
|
|
* coupled cpus, usually the same as cpu_possible_mask if all cpus
|
|
|
|
* are part of the same cluster. The coupled_cpus mask must be
|
|
|
|
* set in the struct cpuidle_device for each cpu.
|
|
|
|
*
|
|
|
|
* Set struct cpuidle_device.safe_state to a state that is not a
|
|
|
|
* coupled state. This is usually WFI.
|
|
|
|
*
|
|
|
|
* Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
|
|
|
|
* state that affects multiple cpus.
|
|
|
|
*
|
|
|
|
* Provide a struct cpuidle_state.enter function for each state
|
|
|
|
* that affects multiple cpus. This function is guaranteed to be
|
|
|
|
* called on all cpus at approximately the same time. The driver
|
|
|
|
* should ensure that the cpus all abort together if any cpu tries
|
|
|
|
* to abort once the function is called. The function should return
|
|
|
|
* with interrupts still disabled.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/**
|
|
|
|
* struct cpuidle_coupled - data for set of cpus that share a coupled idle state
|
|
|
|
* @coupled_cpus: mask of cpus that are part of the coupled set
|
|
|
|
* @requested_state: array of requested states for cpus in the coupled set
|
|
|
|
* @ready_waiting_counts: combined count of cpus in ready or waiting loops
|
2020-01-20 21:24:08 +08:00
|
|
|
* @abort_barrier: synchronisation point for abort cases
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
* @online_count: count of cpus that are online
|
|
|
|
* @refcnt: reference count of cpuidle devices that are using this struct
|
|
|
|
* @prevent: flag to prevent coupled idle while a cpu is hotplugging
|
|
|
|
*/
|
|
|
|
struct cpuidle_coupled {
|
|
|
|
cpumask_t coupled_cpus;
|
|
|
|
int requested_state[NR_CPUS];
|
|
|
|
atomic_t ready_waiting_counts;
|
2013-08-29 09:41:47 +08:00
|
|
|
atomic_t abort_barrier;
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
int online_count;
|
|
|
|
int refcnt;
|
|
|
|
int prevent;
|
|
|
|
};
|
|
|
|
|
|
|
|
#define WAITING_BITS 16
|
|
|
|
#define MAX_WAITING_CPUS (1 << WAITING_BITS)
|
|
|
|
#define WAITING_MASK (MAX_WAITING_CPUS - 1)
|
|
|
|
#define READY_MASK (~WAITING_MASK)
|
|
|
|
|
|
|
|
#define CPUIDLE_COUPLED_NOT_IDLE (-1)
|
|
|
|
|
smp: Avoid using two cache lines for struct call_single_data
struct call_single_data is used in IPIs to transfer information between
CPUs. Its size is bigger than sizeof(unsigned long) and less than
cache line size. Currently it is not allocated with any explicit alignment
requirements. This makes it possible for allocated call_single_data to
cross two cache lines, which results in double the number of the cache lines
that need to be transferred among CPUs.
This can be fixed by requiring call_single_data to be aligned with the
size of call_single_data. Currently the size of call_single_data is the
power of 2. If we add new fields to call_single_data, we may need to
add padding to make sure the size of new definition is the power of 2
as well.
Fortunately, this is enforced by GCC, which will report bad sizes.
To set alignment requirements of call_single_data to the size of
call_single_data, a struct definition and a typedef is used.
To test the effect of the patch, I used the vm-scalability multiple
thread swap test case (swap-w-seq-mt). The test will create multiple
threads and each thread will eat memory until all RAM and part of swap
is used, so that huge number of IPIs are triggered when unmapping
memory. In the test, the throughput of memory writing improves ~5%
compared with misaligned call_single_data, because of faster IPIs.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Huang, Ying <ying.huang@intel.com>
[ Add call_single_data_t and align with size of call_single_data. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/87bmnqd6lz.fsf@yhuang-mobile.sh.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-08 12:30:00 +08:00
|
|
|
static DEFINE_PER_CPU(call_single_data_t, cpuidle_coupled_poke_cb);
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
|
|
|
|
/*
|
2013-08-29 09:41:47 +08:00
|
|
|
* The cpuidle_coupled_poke_pending mask is used to avoid calling
|
smp: Avoid using two cache lines for struct call_single_data
struct call_single_data is used in IPIs to transfer information between
CPUs. Its size is bigger than sizeof(unsigned long) and less than
cache line size. Currently it is not allocated with any explicit alignment
requirements. This makes it possible for allocated call_single_data to
cross two cache lines, which results in double the number of the cache lines
that need to be transferred among CPUs.
This can be fixed by requiring call_single_data to be aligned with the
size of call_single_data. Currently the size of call_single_data is the
power of 2. If we add new fields to call_single_data, we may need to
add padding to make sure the size of new definition is the power of 2
as well.
Fortunately, this is enforced by GCC, which will report bad sizes.
To set alignment requirements of call_single_data to the size of
call_single_data, a struct definition and a typedef is used.
To test the effect of the patch, I used the vm-scalability multiple
thread swap test case (swap-w-seq-mt). The test will create multiple
threads and each thread will eat memory until all RAM and part of swap
is used, so that huge number of IPIs are triggered when unmapping
memory. In the test, the throughput of memory writing improves ~5%
compared with misaligned call_single_data, because of faster IPIs.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Huang, Ying <ying.huang@intel.com>
[ Add call_single_data_t and align with size of call_single_data. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/87bmnqd6lz.fsf@yhuang-mobile.sh.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-08 12:30:00 +08:00
|
|
|
* __smp_call_function_single with the per cpu call_single_data_t struct already
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
* in use. This prevents a deadlock where two cpus are waiting for each others
|
smp: Avoid using two cache lines for struct call_single_data
struct call_single_data is used in IPIs to transfer information between
CPUs. Its size is bigger than sizeof(unsigned long) and less than
cache line size. Currently it is not allocated with any explicit alignment
requirements. This makes it possible for allocated call_single_data to
cross two cache lines, which results in double the number of the cache lines
that need to be transferred among CPUs.
This can be fixed by requiring call_single_data to be aligned with the
size of call_single_data. Currently the size of call_single_data is the
power of 2. If we add new fields to call_single_data, we may need to
add padding to make sure the size of new definition is the power of 2
as well.
Fortunately, this is enforced by GCC, which will report bad sizes.
To set alignment requirements of call_single_data to the size of
call_single_data, a struct definition and a typedef is used.
To test the effect of the patch, I used the vm-scalability multiple
thread swap test case (swap-w-seq-mt). The test will create multiple
threads and each thread will eat memory until all RAM and part of swap
is used, so that huge number of IPIs are triggered when unmapping
memory. In the test, the throughput of memory writing improves ~5%
compared with misaligned call_single_data, because of faster IPIs.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Huang, Ying <ying.huang@intel.com>
[ Add call_single_data_t and align with size of call_single_data. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/87bmnqd6lz.fsf@yhuang-mobile.sh.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-08 12:30:00 +08:00
|
|
|
* call_single_data_t struct to be available
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
*/
|
2013-08-29 09:41:47 +08:00
|
|
|
static cpumask_t cpuidle_coupled_poke_pending;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The cpuidle_coupled_poked mask is used to ensure that each cpu has been poked
|
|
|
|
* once to minimize entering the ready loop with a poke pending, which would
|
|
|
|
* require aborting and retrying.
|
|
|
|
*/
|
|
|
|
static cpumask_t cpuidle_coupled_poked;
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
|
2012-05-08 08:57:42 +08:00
|
|
|
/**
|
|
|
|
* cpuidle_coupled_parallel_barrier - synchronize all online coupled cpus
|
|
|
|
* @dev: cpuidle_device of the calling cpu
|
|
|
|
* @a: atomic variable to hold the barrier
|
|
|
|
*
|
|
|
|
* No caller to this function will return from this function until all online
|
|
|
|
* cpus in the same coupled group have called this function. Once any caller
|
|
|
|
* has returned from this function, the barrier is immediately available for
|
|
|
|
* reuse.
|
|
|
|
*
|
2013-10-03 23:56:41 +08:00
|
|
|
* The atomic variable must be initialized to 0 before any cpu calls
|
2012-05-08 08:57:42 +08:00
|
|
|
* this function, will be reset to 0 before any cpu returns from this function.
|
|
|
|
*
|
|
|
|
* Must only be called from within a coupled idle state handler
|
|
|
|
* (state.enter when state.flags has CPUIDLE_FLAG_COUPLED set).
|
|
|
|
*
|
|
|
|
* Provides full smp barrier semantics before and after calling.
|
|
|
|
*/
|
|
|
|
void cpuidle_coupled_parallel_barrier(struct cpuidle_device *dev, atomic_t *a)
|
|
|
|
{
|
|
|
|
int n = dev->coupled->online_count;
|
|
|
|
|
2014-03-18 01:06:10 +08:00
|
|
|
smp_mb__before_atomic();
|
2012-05-08 08:57:42 +08:00
|
|
|
atomic_inc(a);
|
|
|
|
|
|
|
|
while (atomic_read(a) < n)
|
|
|
|
cpu_relax();
|
|
|
|
|
|
|
|
if (atomic_inc_return(a) == n * 2) {
|
|
|
|
atomic_set(a, 0);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
while (atomic_read(a) > n)
|
|
|
|
cpu_relax();
|
|
|
|
}
|
|
|
|
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
/**
|
|
|
|
* cpuidle_state_is_coupled - check if a state is part of a coupled set
|
|
|
|
* @drv: struct cpuidle_driver for the platform
|
|
|
|
* @state: index of the target state in drv->states
|
|
|
|
*
|
|
|
|
* Returns true if the target state is coupled with cpus besides this one
|
|
|
|
*/
|
2015-08-04 13:48:56 +08:00
|
|
|
bool cpuidle_state_is_coupled(struct cpuidle_driver *drv, int state)
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
{
|
|
|
|
return drv->states[state].flags & CPUIDLE_FLAG_COUPLED;
|
|
|
|
}
|
|
|
|
|
2015-08-31 11:34:05 +08:00
|
|
|
/**
|
|
|
|
* cpuidle_coupled_state_verify - check if the coupled states are correctly set.
|
|
|
|
* @drv: struct cpuidle_driver for the platform
|
|
|
|
*
|
|
|
|
* Returns 0 for valid state values, a negative error code otherwise:
|
|
|
|
* * -EINVAL if any coupled state(safe_state_index) is wrongly set.
|
|
|
|
*/
|
|
|
|
int cpuidle_coupled_state_verify(struct cpuidle_driver *drv)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = drv->state_count - 1; i >= 0; i--) {
|
|
|
|
if (cpuidle_state_is_coupled(drv, i) &&
|
|
|
|
(drv->safe_state_index == i ||
|
|
|
|
drv->safe_state_index < 0 ||
|
|
|
|
drv->safe_state_index >= drv->state_count))
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
/**
|
|
|
|
* cpuidle_coupled_set_ready - mark a cpu as ready
|
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*/
|
|
|
|
static inline void cpuidle_coupled_set_ready(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
atomic_add(MAX_WAITING_CPUS, &coupled->ready_waiting_counts);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_set_not_ready - mark a cpu as not ready
|
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Decrements the ready counter, unless the ready (and thus the waiting) counter
|
|
|
|
* is equal to the number of online cpus. Prevents a race where one cpu
|
|
|
|
* decrements the waiting counter and then re-increments it just before another
|
|
|
|
* cpu has decremented its ready counter, leading to the ready counter going
|
|
|
|
* down from the number of online cpus without going through the coupled idle
|
|
|
|
* state.
|
|
|
|
*
|
|
|
|
* Returns 0 if the counter was decremented successfully, -EINVAL if the ready
|
|
|
|
* counter was equal to the number of online cpus.
|
|
|
|
*/
|
|
|
|
static
|
|
|
|
inline int cpuidle_coupled_set_not_ready(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int all;
|
|
|
|
int ret;
|
|
|
|
|
2012-12-18 20:52:54 +08:00
|
|
|
all = coupled->online_count | (coupled->online_count << WAITING_BITS);
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
ret = atomic_add_unless(&coupled->ready_waiting_counts,
|
|
|
|
-MAX_WAITING_CPUS, all);
|
|
|
|
|
|
|
|
return ret ? 0 : -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_no_cpus_ready - check if no cpus in a coupled set are ready
|
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Returns true if all of the cpus in a coupled set are out of the ready loop.
|
|
|
|
*/
|
|
|
|
static inline int cpuidle_coupled_no_cpus_ready(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int r = atomic_read(&coupled->ready_waiting_counts) >> WAITING_BITS;
|
|
|
|
return r == 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_cpus_ready - check if all cpus in a coupled set are ready
|
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Returns true if all cpus coupled to this target state are in the ready loop
|
|
|
|
*/
|
|
|
|
static inline bool cpuidle_coupled_cpus_ready(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int r = atomic_read(&coupled->ready_waiting_counts) >> WAITING_BITS;
|
|
|
|
return r == coupled->online_count;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_cpus_waiting - check if all cpus in a coupled set are waiting
|
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Returns true if all cpus coupled to this target state are in the wait loop
|
|
|
|
*/
|
|
|
|
static inline bool cpuidle_coupled_cpus_waiting(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int w = atomic_read(&coupled->ready_waiting_counts) & WAITING_MASK;
|
|
|
|
return w == coupled->online_count;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_no_cpus_waiting - check if no cpus in coupled set are waiting
|
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Returns true if all of the cpus in a coupled set are out of the waiting loop.
|
|
|
|
*/
|
|
|
|
static inline int cpuidle_coupled_no_cpus_waiting(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int w = atomic_read(&coupled->ready_waiting_counts) & WAITING_MASK;
|
|
|
|
return w == 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_get_state - determine the deepest idle state
|
|
|
|
* @dev: struct cpuidle_device for this cpu
|
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Returns the deepest idle state that all coupled cpus can enter
|
|
|
|
*/
|
|
|
|
static inline int cpuidle_coupled_get_state(struct cpuidle_device *dev,
|
|
|
|
struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
int state = INT_MAX;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Read barrier ensures that read of requested_state is ordered after
|
|
|
|
* reads of ready_count. Matches the write barriers
|
|
|
|
* cpuidle_set_state_waiting.
|
|
|
|
*/
|
|
|
|
smp_rmb();
|
|
|
|
|
2015-03-05 08:19:16 +08:00
|
|
|
for_each_cpu(i, &coupled->coupled_cpus)
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
if (cpu_online(i) && coupled->requested_state[i] < state)
|
|
|
|
state = coupled->requested_state[i];
|
|
|
|
|
|
|
|
return state;
|
|
|
|
}
|
|
|
|
|
2013-08-29 09:41:47 +08:00
|
|
|
static void cpuidle_coupled_handle_poke(void *info)
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
{
|
|
|
|
int cpu = (unsigned long)info;
|
2013-08-29 09:41:47 +08:00
|
|
|
cpumask_set_cpu(cpu, &cpuidle_coupled_poked);
|
|
|
|
cpumask_clear_cpu(cpu, &cpuidle_coupled_poke_pending);
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_poke - wake up a cpu that may be waiting
|
|
|
|
* @cpu: target cpu
|
|
|
|
*
|
|
|
|
* Ensures that the target cpu exits it's waiting idle state (if it is in it)
|
|
|
|
* and will see updates to waiting_count before it re-enters it's waiting idle
|
|
|
|
* state.
|
|
|
|
*
|
|
|
|
* If cpuidle_coupled_poked_mask is already set for the target cpu, that cpu
|
|
|
|
* either has or will soon have a pending IPI that will wake it out of idle,
|
|
|
|
* or it is currently processing the IPI and is not in idle.
|
|
|
|
*/
|
|
|
|
static void cpuidle_coupled_poke(int cpu)
|
|
|
|
{
|
smp: Avoid using two cache lines for struct call_single_data
struct call_single_data is used in IPIs to transfer information between
CPUs. Its size is bigger than sizeof(unsigned long) and less than
cache line size. Currently it is not allocated with any explicit alignment
requirements. This makes it possible for allocated call_single_data to
cross two cache lines, which results in double the number of the cache lines
that need to be transferred among CPUs.
This can be fixed by requiring call_single_data to be aligned with the
size of call_single_data. Currently the size of call_single_data is the
power of 2. If we add new fields to call_single_data, we may need to
add padding to make sure the size of new definition is the power of 2
as well.
Fortunately, this is enforced by GCC, which will report bad sizes.
To set alignment requirements of call_single_data to the size of
call_single_data, a struct definition and a typedef is used.
To test the effect of the patch, I used the vm-scalability multiple
thread swap test case (swap-w-seq-mt). The test will create multiple
threads and each thread will eat memory until all RAM and part of swap
is used, so that huge number of IPIs are triggered when unmapping
memory. In the test, the throughput of memory writing improves ~5%
compared with misaligned call_single_data, because of faster IPIs.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Huang, Ying <ying.huang@intel.com>
[ Add call_single_data_t and align with size of call_single_data. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/87bmnqd6lz.fsf@yhuang-mobile.sh.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-08 12:30:00 +08:00
|
|
|
call_single_data_t *csd = &per_cpu(cpuidle_coupled_poke_cb, cpu);
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
|
2013-08-29 09:41:47 +08:00
|
|
|
if (!cpumask_test_and_set_cpu(cpu, &cpuidle_coupled_poke_pending))
|
2014-02-24 23:40:02 +08:00
|
|
|
smp_call_function_single_async(cpu, csd);
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_poke_others - wake up all other cpus that may be waiting
|
2020-01-20 21:24:08 +08:00
|
|
|
* @this_cpu: target cpu
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Calls cpuidle_coupled_poke on all other online cpus.
|
|
|
|
*/
|
|
|
|
static void cpuidle_coupled_poke_others(int this_cpu,
|
|
|
|
struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int cpu;
|
|
|
|
|
2015-03-05 08:19:16 +08:00
|
|
|
for_each_cpu(cpu, &coupled->coupled_cpus)
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
if (cpu != this_cpu && cpu_online(cpu))
|
|
|
|
cpuidle_coupled_poke(cpu);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_set_waiting - mark this cpu as in the wait loop
|
2020-01-20 21:24:08 +08:00
|
|
|
* @cpu: target cpu
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
* @next_state: the index in drv->states of the requested state for this cpu
|
|
|
|
*
|
2013-08-29 09:41:47 +08:00
|
|
|
* Updates the requested idle state for the specified cpuidle device.
|
|
|
|
* Returns the number of waiting cpus.
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
*/
|
2013-08-29 09:41:47 +08:00
|
|
|
static int cpuidle_coupled_set_waiting(int cpu,
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
struct cpuidle_coupled *coupled, int next_state)
|
|
|
|
{
|
|
|
|
coupled->requested_state[cpu] = next_state;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The atomic_inc_return provides a write barrier to order the write
|
|
|
|
* to requested_state with the later write that increments ready_count.
|
|
|
|
*/
|
2013-08-29 09:41:47 +08:00
|
|
|
return atomic_inc_return(&coupled->ready_waiting_counts) & WAITING_MASK;
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_set_not_waiting - mark this cpu as leaving the wait loop
|
2020-01-20 21:24:08 +08:00
|
|
|
* @cpu: target cpu
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Removes the requested idle state for the specified cpuidle device.
|
|
|
|
*/
|
|
|
|
static void cpuidle_coupled_set_not_waiting(int cpu,
|
|
|
|
struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
/*
|
|
|
|
* Decrementing waiting count can race with incrementing it in
|
|
|
|
* cpuidle_coupled_set_waiting, but that's OK. Worst case, some
|
|
|
|
* cpus will increment ready_count and then spin until they
|
|
|
|
* notice that this cpu has cleared it's requested_state.
|
|
|
|
*/
|
|
|
|
atomic_dec(&coupled->ready_waiting_counts);
|
|
|
|
|
|
|
|
coupled->requested_state[cpu] = CPUIDLE_COUPLED_NOT_IDLE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_set_done - mark this cpu as leaving the ready loop
|
|
|
|
* @cpu: the current cpu
|
|
|
|
* @coupled: the struct coupled that contains the current cpu
|
|
|
|
*
|
|
|
|
* Marks this cpu as no longer in the ready and waiting loops. Decrements
|
|
|
|
* the waiting count first to prevent another cpu looping back in and seeing
|
|
|
|
* this cpu as waiting just before it exits idle.
|
|
|
|
*/
|
|
|
|
static void cpuidle_coupled_set_done(int cpu, struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
cpuidle_coupled_set_not_waiting(cpu, coupled);
|
|
|
|
atomic_sub(MAX_WAITING_CPUS, &coupled->ready_waiting_counts);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_clear_pokes - spin until the poke interrupt is processed
|
2020-01-20 21:24:08 +08:00
|
|
|
* @cpu: this cpu
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
*
|
|
|
|
* Turns on interrupts and spins until any outstanding poke interrupts have
|
|
|
|
* been processed and the poke bit has been cleared.
|
|
|
|
*
|
|
|
|
* Other interrupts may also be processed while interrupts are enabled, so
|
2013-08-24 03:45:12 +08:00
|
|
|
* need_resched() must be tested after this function returns to make sure
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
* the interrupt didn't schedule work that should take the cpu out of idle.
|
|
|
|
*
|
2013-08-24 03:45:12 +08:00
|
|
|
* Returns 0 if no poke was pending, 1 if a poke was cleared.
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
*/
|
|
|
|
static int cpuidle_coupled_clear_pokes(int cpu)
|
|
|
|
{
|
2013-08-24 03:45:12 +08:00
|
|
|
if (!cpumask_test_cpu(cpu, &cpuidle_coupled_poke_pending))
|
|
|
|
return 0;
|
|
|
|
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
local_irq_enable();
|
2013-08-29 09:41:47 +08:00
|
|
|
while (cpumask_test_cpu(cpu, &cpuidle_coupled_poke_pending))
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
cpu_relax();
|
|
|
|
local_irq_disable();
|
|
|
|
|
2013-08-24 03:45:12 +08:00
|
|
|
return 1;
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
}
|
|
|
|
|
2013-08-29 09:41:47 +08:00
|
|
|
static bool cpuidle_coupled_any_pokes_pending(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
cpumask_t cpus;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
cpumask_and(&cpus, cpu_online_mask, &coupled->coupled_cpus);
|
|
|
|
ret = cpumask_and(&cpus, &cpuidle_coupled_poke_pending, &cpus);
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
/**
|
|
|
|
* cpuidle_enter_state_coupled - attempt to enter a state with coupled cpus
|
|
|
|
* @dev: struct cpuidle_device for the current cpu
|
|
|
|
* @drv: struct cpuidle_driver for the platform
|
|
|
|
* @next_state: index of the requested state in drv->states
|
|
|
|
*
|
|
|
|
* Coordinate with coupled cpus to enter the target state. This is a two
|
|
|
|
* stage process. In the first stage, the cpus are operating independently,
|
|
|
|
* and may call into cpuidle_enter_state_coupled at completely different times.
|
|
|
|
* To save as much power as possible, the first cpus to call this function will
|
|
|
|
* go to an intermediate state (the cpuidle_device's safe state), and wait for
|
|
|
|
* all the other cpus to call this function. Once all coupled cpus are idle,
|
|
|
|
* the second stage will start. Each coupled cpu will spin until all cpus have
|
|
|
|
* guaranteed that they will call the target_state.
|
|
|
|
*
|
|
|
|
* This function must be called with interrupts disabled. It may enable
|
|
|
|
* interrupts while preparing for idle, and it will always return with
|
|
|
|
* interrupts enabled.
|
|
|
|
*/
|
|
|
|
int cpuidle_enter_state_coupled(struct cpuidle_device *dev,
|
|
|
|
struct cpuidle_driver *drv, int next_state)
|
|
|
|
{
|
|
|
|
int entered_state = -1;
|
|
|
|
struct cpuidle_coupled *coupled = dev->coupled;
|
2013-08-29 09:41:47 +08:00
|
|
|
int w;
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
|
|
|
|
if (!coupled)
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
while (coupled->prevent) {
|
2013-08-24 03:45:12 +08:00
|
|
|
cpuidle_coupled_clear_pokes(dev->cpu);
|
|
|
|
if (need_resched()) {
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
local_irq_enable();
|
|
|
|
return entered_state;
|
|
|
|
}
|
|
|
|
entered_state = cpuidle_enter_state(dev, drv,
|
2015-08-04 13:48:55 +08:00
|
|
|
drv->safe_state_index);
|
2013-08-24 03:45:10 +08:00
|
|
|
local_irq_disable();
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Read barrier ensures online_count is read after prevent is cleared */
|
|
|
|
smp_rmb();
|
|
|
|
|
2013-08-29 09:41:47 +08:00
|
|
|
reset:
|
|
|
|
cpumask_clear_cpu(dev->cpu, &cpuidle_coupled_poked);
|
|
|
|
|
|
|
|
w = cpuidle_coupled_set_waiting(dev->cpu, coupled, next_state);
|
|
|
|
/*
|
|
|
|
* If this is the last cpu to enter the waiting state, poke
|
|
|
|
* all the other cpus out of their waiting state so they can
|
|
|
|
* enter a deeper state. This can race with one of the cpus
|
|
|
|
* exiting the waiting state due to an interrupt and
|
|
|
|
* decrementing waiting_count, see comment below.
|
|
|
|
*/
|
|
|
|
if (w == coupled->online_count) {
|
|
|
|
cpumask_set_cpu(dev->cpu, &cpuidle_coupled_poked);
|
|
|
|
cpuidle_coupled_poke_others(dev->cpu, coupled);
|
|
|
|
}
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
|
|
|
|
retry:
|
|
|
|
/*
|
|
|
|
* Wait for all coupled cpus to be idle, using the deepest state
|
2013-08-29 09:41:47 +08:00
|
|
|
* allowed for a single cpu. If this was not the poking cpu, wait
|
|
|
|
* for at least one poke before leaving to avoid a race where
|
|
|
|
* two cpus could arrive at the waiting loop at the same time,
|
|
|
|
* but the first of the two to arrive could skip the loop without
|
|
|
|
* processing the pokes from the last to arrive.
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
*/
|
2013-08-29 09:41:47 +08:00
|
|
|
while (!cpuidle_coupled_cpus_waiting(coupled) ||
|
|
|
|
!cpumask_test_cpu(dev->cpu, &cpuidle_coupled_poked)) {
|
2013-08-24 03:45:12 +08:00
|
|
|
if (cpuidle_coupled_clear_pokes(dev->cpu))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (need_resched()) {
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
cpuidle_coupled_set_not_waiting(dev->cpu, coupled);
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (coupled->prevent) {
|
|
|
|
cpuidle_coupled_set_not_waiting(dev->cpu, coupled);
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
entered_state = cpuidle_enter_state(dev, drv,
|
2015-08-04 13:48:55 +08:00
|
|
|
drv->safe_state_index);
|
2013-08-24 03:45:10 +08:00
|
|
|
local_irq_disable();
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
}
|
|
|
|
|
2013-08-24 03:45:12 +08:00
|
|
|
cpuidle_coupled_clear_pokes(dev->cpu);
|
|
|
|
if (need_resched()) {
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
cpuidle_coupled_set_not_waiting(dev->cpu, coupled);
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
2013-08-29 09:41:47 +08:00
|
|
|
/*
|
|
|
|
* Make sure final poke status for this cpu is visible before setting
|
|
|
|
* cpu as ready.
|
|
|
|
*/
|
|
|
|
smp_wmb();
|
|
|
|
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
/*
|
|
|
|
* All coupled cpus are probably idle. There is a small chance that
|
|
|
|
* one of the other cpus just became active. Increment the ready count,
|
|
|
|
* and spin until all coupled cpus have incremented the counter. Once a
|
|
|
|
* cpu has incremented the ready counter, it cannot abort idle and must
|
|
|
|
* spin until either all cpus have incremented the ready counter, or
|
|
|
|
* another cpu leaves idle and decrements the waiting counter.
|
|
|
|
*/
|
|
|
|
|
|
|
|
cpuidle_coupled_set_ready(coupled);
|
|
|
|
while (!cpuidle_coupled_cpus_ready(coupled)) {
|
|
|
|
/* Check if any other cpus bailed out of idle. */
|
|
|
|
if (!cpuidle_coupled_cpus_waiting(coupled))
|
|
|
|
if (!cpuidle_coupled_set_not_ready(coupled))
|
|
|
|
goto retry;
|
|
|
|
|
|
|
|
cpu_relax();
|
|
|
|
}
|
|
|
|
|
2013-08-29 09:41:47 +08:00
|
|
|
/*
|
|
|
|
* Make sure read of all cpus ready is done before reading pending pokes
|
|
|
|
*/
|
|
|
|
smp_rmb();
|
|
|
|
|
|
|
|
/*
|
|
|
|
* There is a small chance that a cpu left and reentered idle after this
|
|
|
|
* cpu saw that all cpus were waiting. The cpu that reentered idle will
|
|
|
|
* have sent this cpu a poke, which will still be pending after the
|
|
|
|
* ready loop. The pending interrupt may be lost by the interrupt
|
|
|
|
* controller when entering the deep idle state. It's not possible to
|
|
|
|
* clear a pending interrupt without turning interrupts on and handling
|
|
|
|
* it, and it's too late to turn on interrupts here, so reset the
|
|
|
|
* coupled idle state of all cpus and retry.
|
|
|
|
*/
|
|
|
|
if (cpuidle_coupled_any_pokes_pending(coupled)) {
|
|
|
|
cpuidle_coupled_set_done(dev->cpu, coupled);
|
|
|
|
/* Wait for all cpus to see the pending pokes */
|
|
|
|
cpuidle_coupled_parallel_barrier(dev, &coupled->abort_barrier);
|
|
|
|
goto reset;
|
|
|
|
}
|
|
|
|
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
/* all cpus have acked the coupled state */
|
|
|
|
next_state = cpuidle_coupled_get_state(dev, coupled);
|
|
|
|
|
|
|
|
entered_state = cpuidle_enter_state(dev, drv, next_state);
|
|
|
|
|
|
|
|
cpuidle_coupled_set_done(dev->cpu, coupled);
|
|
|
|
|
|
|
|
out:
|
|
|
|
/*
|
|
|
|
* Normal cpuidle states are expected to return with irqs enabled.
|
|
|
|
* That leads to an inefficiency where a cpu receiving an interrupt
|
|
|
|
* that brings it out of idle will process that interrupt before
|
|
|
|
* exiting the idle enter function and decrementing ready_count. All
|
|
|
|
* other cpus will need to spin waiting for the cpu that is processing
|
|
|
|
* the interrupt. If the driver returns with interrupts disabled,
|
|
|
|
* all other cpus will loop back into the safe idle state instead of
|
|
|
|
* spinning, saving power.
|
|
|
|
*
|
|
|
|
* Calling local_irq_enable here allows coupled states to return with
|
|
|
|
* interrupts disabled, but won't cause problems for drivers that
|
|
|
|
* exit with interrupts enabled.
|
|
|
|
*/
|
|
|
|
local_irq_enable();
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Wait until all coupled cpus have exited idle. There is no risk that
|
|
|
|
* a cpu exits and re-enters the ready state because this cpu has
|
|
|
|
* already decremented its waiting_count.
|
|
|
|
*/
|
|
|
|
while (!cpuidle_coupled_no_cpus_ready(coupled))
|
|
|
|
cpu_relax();
|
|
|
|
|
|
|
|
return entered_state;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void cpuidle_coupled_update_online_cpus(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
cpumask_t cpus;
|
|
|
|
cpumask_and(&cpus, cpu_online_mask, &coupled->coupled_cpus);
|
|
|
|
coupled->online_count = cpumask_weight(&cpus);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_register_device - register a coupled cpuidle device
|
|
|
|
* @dev: struct cpuidle_device for the current cpu
|
|
|
|
*
|
|
|
|
* Called from cpuidle_register_device to handle coupled idle init. Finds the
|
|
|
|
* cpuidle_coupled struct for this set of coupled cpus, or creates one if none
|
|
|
|
* exists yet.
|
|
|
|
*/
|
|
|
|
int cpuidle_coupled_register_device(struct cpuidle_device *dev)
|
|
|
|
{
|
|
|
|
int cpu;
|
|
|
|
struct cpuidle_device *other_dev;
|
smp: Avoid using two cache lines for struct call_single_data
struct call_single_data is used in IPIs to transfer information between
CPUs. Its size is bigger than sizeof(unsigned long) and less than
cache line size. Currently it is not allocated with any explicit alignment
requirements. This makes it possible for allocated call_single_data to
cross two cache lines, which results in double the number of the cache lines
that need to be transferred among CPUs.
This can be fixed by requiring call_single_data to be aligned with the
size of call_single_data. Currently the size of call_single_data is the
power of 2. If we add new fields to call_single_data, we may need to
add padding to make sure the size of new definition is the power of 2
as well.
Fortunately, this is enforced by GCC, which will report bad sizes.
To set alignment requirements of call_single_data to the size of
call_single_data, a struct definition and a typedef is used.
To test the effect of the patch, I used the vm-scalability multiple
thread swap test case (swap-w-seq-mt). The test will create multiple
threads and each thread will eat memory until all RAM and part of swap
is used, so that huge number of IPIs are triggered when unmapping
memory. In the test, the throughput of memory writing improves ~5%
compared with misaligned call_single_data, because of faster IPIs.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Huang, Ying <ying.huang@intel.com>
[ Add call_single_data_t and align with size of call_single_data. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Aaron Lu <aaron.lu@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/87bmnqd6lz.fsf@yhuang-mobile.sh.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-08-08 12:30:00 +08:00
|
|
|
call_single_data_t *csd;
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
struct cpuidle_coupled *coupled;
|
|
|
|
|
|
|
|
if (cpumask_empty(&dev->coupled_cpus))
|
|
|
|
return 0;
|
|
|
|
|
2015-03-05 08:19:16 +08:00
|
|
|
for_each_cpu(cpu, &dev->coupled_cpus) {
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
other_dev = per_cpu(cpuidle_devices, cpu);
|
|
|
|
if (other_dev && other_dev->coupled) {
|
|
|
|
coupled = other_dev->coupled;
|
|
|
|
goto have_coupled;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* No existing coupled info found, create a new one */
|
|
|
|
coupled = kzalloc(sizeof(struct cpuidle_coupled), GFP_KERNEL);
|
|
|
|
if (!coupled)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
coupled->coupled_cpus = dev->coupled_cpus;
|
|
|
|
|
|
|
|
have_coupled:
|
|
|
|
dev->coupled = coupled;
|
|
|
|
if (WARN_ON(!cpumask_equal(&dev->coupled_cpus, &coupled->coupled_cpus)))
|
|
|
|
coupled->prevent++;
|
|
|
|
|
|
|
|
cpuidle_coupled_update_online_cpus(coupled);
|
|
|
|
|
|
|
|
coupled->refcnt++;
|
|
|
|
|
|
|
|
csd = &per_cpu(cpuidle_coupled_poke_cb, dev->cpu);
|
2020-06-15 17:29:31 +08:00
|
|
|
INIT_CSD(csd, cpuidle_coupled_handle_poke, (void *)(unsigned long)dev->cpu);
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_unregister_device - unregister a coupled cpuidle device
|
|
|
|
* @dev: struct cpuidle_device for the current cpu
|
|
|
|
*
|
|
|
|
* Called from cpuidle_unregister_device to tear down coupled idle. Removes the
|
|
|
|
* cpu from the coupled idle set, and frees the cpuidle_coupled_info struct if
|
|
|
|
* this was the last cpu in the set.
|
|
|
|
*/
|
|
|
|
void cpuidle_coupled_unregister_device(struct cpuidle_device *dev)
|
|
|
|
{
|
|
|
|
struct cpuidle_coupled *coupled = dev->coupled;
|
|
|
|
|
|
|
|
if (cpumask_empty(&dev->coupled_cpus))
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (--coupled->refcnt)
|
|
|
|
kfree(coupled);
|
|
|
|
dev->coupled = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_prevent_idle - prevent cpus from entering a coupled state
|
|
|
|
* @coupled: the struct coupled that contains the cpu that is changing state
|
|
|
|
*
|
|
|
|
* Disables coupled cpuidle on a coupled set of cpus. Used to ensure that
|
|
|
|
* cpu_online_mask doesn't change while cpus are coordinating coupled idle.
|
|
|
|
*/
|
|
|
|
static void cpuidle_coupled_prevent_idle(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int cpu = get_cpu();
|
|
|
|
|
|
|
|
/* Force all cpus out of the waiting loop. */
|
|
|
|
coupled->prevent++;
|
|
|
|
cpuidle_coupled_poke_others(cpu, coupled);
|
|
|
|
put_cpu();
|
|
|
|
while (!cpuidle_coupled_no_cpus_waiting(coupled))
|
|
|
|
cpu_relax();
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* cpuidle_coupled_allow_idle - allows cpus to enter a coupled state
|
|
|
|
* @coupled: the struct coupled that contains the cpu that is changing state
|
|
|
|
*
|
|
|
|
* Enables coupled cpuidle on a coupled set of cpus. Used to ensure that
|
|
|
|
* cpu_online_mask doesn't change while cpus are coordinating coupled idle.
|
|
|
|
*/
|
|
|
|
static void cpuidle_coupled_allow_idle(struct cpuidle_coupled *coupled)
|
|
|
|
{
|
|
|
|
int cpu = get_cpu();
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Write barrier ensures readers see the new online_count when they
|
|
|
|
* see prevent == 0.
|
|
|
|
*/
|
|
|
|
smp_wmb();
|
|
|
|
coupled->prevent--;
|
|
|
|
/* Force cpus out of the prevent loop. */
|
|
|
|
cpuidle_coupled_poke_others(cpu, coupled);
|
|
|
|
put_cpu();
|
|
|
|
}
|
|
|
|
|
2016-08-24 17:14:44 +08:00
|
|
|
static int coupled_cpu_online(unsigned int cpu)
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
{
|
|
|
|
struct cpuidle_device *dev;
|
|
|
|
|
|
|
|
mutex_lock(&cpuidle_lock);
|
|
|
|
|
|
|
|
dev = per_cpu(cpuidle_devices, cpu);
|
2016-08-24 17:14:44 +08:00
|
|
|
if (dev && dev->coupled) {
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
cpuidle_coupled_update_online_cpus(dev->coupled);
|
|
|
|
cpuidle_coupled_allow_idle(dev->coupled);
|
|
|
|
}
|
|
|
|
|
|
|
|
mutex_unlock(&cpuidle_lock);
|
2016-08-24 17:14:44 +08:00
|
|
|
return 0;
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
}
|
|
|
|
|
2016-08-24 17:14:44 +08:00
|
|
|
static int coupled_cpu_up_prepare(unsigned int cpu)
|
|
|
|
{
|
|
|
|
struct cpuidle_device *dev;
|
|
|
|
|
|
|
|
mutex_lock(&cpuidle_lock);
|
|
|
|
|
|
|
|
dev = per_cpu(cpuidle_devices, cpu);
|
|
|
|
if (dev && dev->coupled)
|
|
|
|
cpuidle_coupled_prevent_idle(dev->coupled);
|
|
|
|
|
|
|
|
mutex_unlock(&cpuidle_lock);
|
|
|
|
return 0;
|
|
|
|
}
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
|
|
|
|
static int __init cpuidle_coupled_init(void)
|
|
|
|
{
|
2016-08-24 17:14:44 +08:00
|
|
|
int ret;
|
|
|
|
|
|
|
|
ret = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_COUPLED_PREPARE,
|
|
|
|
"cpuidle/coupled:prepare",
|
|
|
|
coupled_cpu_up_prepare,
|
|
|
|
coupled_cpu_online);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
|
|
|
|
"cpuidle/coupled:online",
|
|
|
|
coupled_cpu_online,
|
|
|
|
coupled_cpu_up_prepare);
|
|
|
|
if (ret < 0)
|
|
|
|
cpuhp_remove_state_nocalls(CPUHP_CPUIDLE_COUPLED_PREPARE);
|
|
|
|
return ret;
|
cpuidle: add support for states that affect multiple cpus
On some ARM SMP SoCs (OMAP4460, Tegra 2, and probably more), the
cpus cannot be independently powered down, either due to
sequencing restrictions (on Tegra 2, cpu 0 must be the last to
power down), or due to HW bugs (on OMAP4460, a cpu powering up
will corrupt the gic state unless the other cpu runs a work
around). Each cpu has a power state that it can enter without
coordinating with the other cpu (usually Wait For Interrupt, or
WFI), and one or more "coupled" power states that affect blocks
shared between the cpus (L2 cache, interrupt controller, and
sometimes the whole SoC). Entering a coupled power state must
be tightly controlled on both cpus.
The easiest solution to implementing coupled cpu power states is
to hotplug all but one cpu whenever possible, usually using a
cpufreq governor that looks at cpu load to determine when to
enable the secondary cpus. This causes problems, as hotplug is an
expensive operation, so the number of hotplug transitions must be
minimized, leading to very slow response to loads, often on the
order of seconds.
This file implements an alternative solution, where each cpu will
wait in the WFI state until all cpus are ready to enter a coupled
state, at which point the coupled state function will be called
on all cpus at approximately the same time.
Once all cpus are ready to enter idle, they are woken by an smp
cross call. At this point, there is a chance that one of the
cpus will find work to do, and choose not to enter idle. A
final pass is needed to guarantee that all cpus will call the
power state enter function at the same time. During this pass,
each cpu will increment the ready counter, and continue once the
ready counter matches the number of online coupled cpus. If any
cpu exits idle, the other cpus will decrement their counter and
retry.
To use coupled cpuidle states, a cpuidle driver must:
Set struct cpuidle_device.coupled_cpus to the mask of all
coupled cpus, usually the same as cpu_possible_mask if all cpus
are part of the same cluster. The coupled_cpus mask must be
set in the struct cpuidle_device for each cpu.
Set struct cpuidle_device.safe_state to a state that is not a
coupled state. This is usually WFI.
Set CPUIDLE_FLAG_COUPLED in struct cpuidle_state.flags for each
state that affects multiple cpus.
Provide a struct cpuidle_state.enter function for each state
that affects multiple cpus. This function is guaranteed to be
called on all cpus at approximately the same time. The driver
should ensure that the cpus all abort together if any cpu tries
to abort once the function is called.
update1:
cpuidle: coupled: fix count of online cpus
online_count was never incremented on boot, and was also counting
cpus that were not part of the coupled set. Fix both issues by
introducting a new function that counts online coupled cpus, and
call it from register as well as the hotplug notifier.
update2:
cpuidle: coupled: fix decrementing ready count
cpuidle_coupled_set_not_ready sometimes refuses to decrement the
ready count in order to prevent a race condition. This makes it
unsuitable for use when finished with idle. Add a new function
cpuidle_coupled_set_done that decrements both the ready count and
waiting count, and call it after idle is complete.
Cc: Amit Kucheria <amit.kucheria@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Trinabh Gupta <g.trinabh@gmail.com>
Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Reviewed-by: Kevin Hilman <khilman@ti.com>
Tested-by: Kevin Hilman <khilman@ti.com>
Signed-off-by: Colin Cross <ccross@android.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Len Brown <len.brown@intel.com>
2012-05-08 08:57:41 +08:00
|
|
|
}
|
|
|
|
core_initcall(cpuidle_coupled_init);
|