CPU Idle Time Management¶
Copyright (c) 2018 Intel Corp., Rafael J. Wysocki <email@example.com>
Modern processors are generally able to enter states in which the execution of a program is suspended and instructions belonging to it are not fetched from memory or executed. Those states are the idle states of the processor.
Since part of the processor hardware is not used in idle states, entering them generally allows power drawn by the processor to be reduced and, in consequence, it is an opportunity to save energy.
CPU idle time management is an energy-efficiency feature concerned about using the idle states of processors for this purpose.
CPU idle time management operates on CPUs as seen by the CPU scheduler (that is the part of the kernel responsible for the distribution of computational work in the system). In its view, CPUs are logical units. That is, they need not be separate physical entities and may just be interfaces appearing to software as individual single-core processors. In other words, a CPU is an entity which appears to be fetching instructions that belong to one sequence (program) from memory and executing them, but it need not work this way physically. Generally, three different cases can be consider here.
First, if the whole processor can only follow one sequence of instructions (one program) at a time, it is a CPU. In that case, if the hardware is asked to enter an idle state, that applies to the processor as a whole.
Second, if the processor is multi-core, each core in it is able to follow at least one program at a time. The cores need not be entirely independent of each other (for example, they may share caches), but still most of the time they work physically in parallel with each other, so if each of them executes only one program, those programs run mostly independently of each other at the same time. The entire cores are CPUs in that case and if the hardware is asked to enter an idle state, that applies to the core that asked for it in the first place, but it also may apply to a larger unit (say a “package” or a “cluster”) that the core belongs to (in fact, it may apply to an entire hierarchy of larger units containing the core). Namely, if all of the cores in the larger unit except for one have been put into idle states at the “core level” and the remaining core asks the processor to enter an idle state, that may trigger it to put the whole larger unit into an idle state which also will affect the other cores in that unit.
Finally, each core in a multi-core processor may be able to follow more than one program in the same time frame (that is, each core may be able to fetch instructions from multiple locations in memory and execute them in the same time frame, but not necessarily entirely in parallel with each other). In that case the cores present themselves to software as “bundles” each consisting of multiple individual single-core “processors”, referred to as hardware threads (or hyper-threads specifically on Intel hardware), that each can follow one sequence of instructions. Then, the hardware threads are CPUs from the CPU idle time management perspective and if the processor is asked to enter an idle state by one of them, the hardware thread (or CPU) that asked for it is stopped, but nothing more happens, unless all of the other hardware threads within the same core also have asked the processor to enter an idle state. In that situation, the core may be put into an idle state individually or a larger unit containing it may be put into an idle state as a whole (if the other cores within the larger unit are in idle states already).
Logical CPUs, simply referred to as “CPUs” in what follows, are regarded as idle by the Linux kernel when there are no tasks to run on them except for the special “idle” task.
Tasks are the CPU scheduler’s representation of work. Each task consists of a sequence of instructions to execute, or code, data to be manipulated while running that code, and some context information that needs to be loaded into the processor every time the task’s code is run by a CPU. The CPU scheduler distributes work by assigning tasks to run to the CPUs present in the system.
Tasks can be in various states. In particular, they are runnable if there are no specific conditions preventing their code from being run by a CPU as long as there is a CPU available for that (for example, they are not waiting for any events to occur or similar). When a task becomes runnable, the CPU scheduler assigns it to one of the available CPUs to run and if there are no more runnable tasks assigned to it, the CPU will load the given task’s context and run its code (from the instruction following the last one executed so far, possibly by another CPU). [If there are multiple runnable tasks assigned to one CPU simultaneously, they will be subject to prioritization and time sharing in order to allow them to make some progress over time.]
The special “idle” task becomes runnable if there are no other runnable tasks assigned to the given CPU and the CPU is then regarded as idle. In other words, in Linux idle CPUs run the code of the “idle” task called the idle loop. That code may cause the processor to be put into one of its idle states, if they are supported, in order to save energy, but if the processor does not support any idle states, or there is not enough time to spend in an idle state before the next wakeup event, or there are strict latency constraints preventing any of the available idle states from being used, the CPU will simply execute more or less useless instructions in a loop until it is assigned a new task to run.
The Idle Loop¶
The idle loop code takes two major steps in every iteration of it. First, it
calls into a code module referred to as the governor that belongs to the CPU
idle time management subsystem called
CPUIdle to select an idle state for
the CPU to ask the hardware to enter. Second, it invokes another code module
CPUIdle subsystem, called the driver, to actually ask the
processor hardware to enter the idle state selected by the governor.
The role of the governor is to find an idle state most suitable for the
conditions at hand. For this purpose, idle states that the hardware can be
asked to enter by logical CPUs are represented in an abstract way independent of
the platform or the processor architecture and organized in a one-dimensional
(linear) array. That array has to be prepared and supplied by the
driver matching the platform the kernel is running on at the initialization
time. This allows
CPUIdle governors to be independent of the underlying
hardware and to work with any platforms that the Linux kernel can run on.
Each idle state present in that array is characterized by two parameters to be taken into account by the governor, the target residency and the (worst-case) exit latency. The target residency is the minimum time the hardware must spend in the given state, including the time needed to enter it (which may be substantial), in order to save more energy than it would save by entering one of the shallower idle states instead. [The “depth” of an idle state roughly corresponds to the power drawn by the processor in that state.] The exit latency, in turn, is the maximum time it will take a CPU asking the processor hardware to enter an idle state to start executing the first instruction after a wakeup from that state. Note that in general the exit latency also must cover the time needed to enter the given state in case the wakeup occurs when the hardware is entering it and it must be entered completely to be exited in an ordered manner.
There are two types of information that can influence the governor’s decisions.
First of all, the governor knows the time until the closest timer event. That
time is known exactly, because the kernel programs timers and it knows exactly
when they will trigger, and it is the maximum time the hardware that the given
CPU depends on can spend in an idle state, including the time necessary to enter
and exit it. However, the CPU may be woken up by a non-timer event at any time
(in particular, before the closest timer triggers) and it generally is not known
when that may happen. The governor can only see how much time the CPU actually
was idle after it has been woken up (that time will be referred to as the idle
duration from now on) and it can use that information somehow along with the
time until the closest timer to estimate the idle duration in future. How the
governor uses that information depends on what algorithm is implemented by it
and that is the primary reason for having more than one governor in the
There are two
CPUIdle governors available,
of them is used depends on the configuration of the kernel and in particular on
whether or not the scheduler tick can be stopped by the idle
loop. It is possible to change the governor at run time
cpuidle_sysfs_switch command line parameter has been passed to the
kernel, but that is not safe in general, so it should not be done on production
systems (that may change in the future, though). The name of the
governor currently used by the kernel can be read from the
cpuidle_sysfs_switch is present in the kernel command line) file under
CPUIdle driver is used, on the other hand, usually depends on the
platform the kernel is running on, but there are platforms with more than one
matching driver. For example, there are two drivers that can work with the
majority of Intel platforms,
acpi_idle, one with
hardcoded idle states information and the other able to read that information
from the system’s ACPI tables, respectively. Still, even in those cases, the
driver chosen at the system initialization time cannot be replaced later, so the
decision on which one of them to use has to be made early (on Intel platforms
acpi_idle driver will be used if
intel_idle is disabled for some
reason or if it does not recognize the processor). The name of the
driver currently used by the kernel can be read from the
Idle CPUs and The Scheduler Tick¶
The scheduler tick is a timer that triggers periodically in order to implement the time sharing strategy of the CPU scheduler. Of course, if there are multiple runnable tasks assigned to one CPU at the same time, the only way to allow them to make reasonable progress in a given time frame is to make them share the available CPU time. Namely, in rough approximation, each task is given a slice of the CPU time to run its code, subject to the scheduling class, prioritization and so on and when that time slice is used up, the CPU should be switched over to running (the code of) another task. The currently running task may not want to give the CPU away voluntarily, however, and the scheduler tick is there to make the switch happen regardless. That is not the only role of the tick, but it is the primary reason for using it.
The scheduler tick is problematic from the CPU idle time management perspective, because it triggers periodically and relatively often (depending on the kernel configuration, the length of the tick period is between 1 ms and 10 ms). Thus, if the tick is allowed to trigger on idle CPUs, it will not make sense for them to ask the hardware to enter idle states with target residencies above the tick period length. Moreover, in that case the idle duration of any CPU will never exceed the tick period length and the energy used for entering and exiting idle states due to the tick wakeups on idle CPUs will be wasted.
Fortunately, it is not really necessary to allow the tick to trigger on idle CPUs, because (by definition) they have no tasks to run except for the special “idle” one. In other words, from the CPU scheduler perspective, the only user of the CPU time on them is the idle loop. Since the time of an idle CPU need not be shared between multiple runnable tasks, the primary reason for using the tick goes away if the given CPU is idle. Consequently, it is possible to stop the scheduler tick entirely on idle CPUs in principle, even though that may not always be worth the effort.
Whether or not it makes sense to stop the scheduler tick in the idle loop depends on what is expected by the governor. First, if there is another (non-tick) timer due to trigger within the tick range, stopping the tick clearly would be a waste of time, even though the timer hardware may not need to be reprogrammed in that case. Second, if the governor is expecting a non-timer wakeup within the tick range, stopping the tick is not necessary and it may even be harmful. Namely, in that case the governor will select an idle state with the target residency within the time until the expected wakeup, so that state is going to be relatively shallow. The governor really cannot select a deep idle state then, as that would contradict its own expectation of a wakeup in short order. Now, if the wakeup really occurs shortly, stopping the tick would be a waste of time and in this case the timer hardware would need to be reprogrammed, which is expensive. On the other hand, if the tick is stopped and the wakeup does not occur any time soon, the hardware may spend indefinite amount of time in the shallow idle state selected by the governor, which will be a waste of energy. Hence, if the governor is expecting a wakeup of any kind within the tick range, it is better to allow the tick trigger. Otherwise, however, the governor will select a relatively deep idle state, so the tick should be stopped so that it does not wake up the CPU too early.
In any case, the governor knows what it is expecting and the decision on whether or not to stop the scheduler tick belongs to it. Still, if the tick has been stopped already (in one of the previous iterations of the loop), it is better to leave it as is and the governor needs to take that into account.
The kernel can be configured to disable stopping the scheduler tick in the idle
loop altogether. That can be done through the build-time configuration of it
(by unsetting the
CONFIG_NO_HZ_IDLE configuration option) or by passing
nohz=off to it in the command line. In both cases, as the stopping of the
scheduler tick is disabled, the governor’s decisions regarding it are simply
ignored by the idle loop code and the tick is never stopped.
The systems that run kernels configured to allow the scheduler tick to be
stopped on idle CPUs are referred to as tickless systems and they are
generally regarded as more energy-efficient than the systems running kernels in
which the tick cannot be stopped. If the given system is tickless, it will use
menu governor by default and if it is not tickless, the default
CPUIdle governor on it will be
Representation of Idle States¶
For the CPU idle time management purposes all of the physical idle states
supported by the processor have to be represented as a one-dimensional array of
struct cpuidle_state objects each allowing an individual (logical) CPU to ask
the processor hardware to enter an idle state of certain properties. If there
is a hierarchy of units in the processor, one
struct cpuidle_state object can
cover a combination of idle states supported by the units at different levels of
the hierarchy. In that case, the target residency and exit latency parameters
of it, must reflect the properties of the idle state at the
deepest level (i.e. the idle state of the unit containing all of the other
For example, take a processor with two cores in a larger unit referred to as
a “module” and suppose that asking the hardware to enter a specific idle state
(say “X”) at the “core” level by one core will trigger the module to try to
enter a specific idle state of its own (say “MX”) if the other core is in idle
state “X” already. In other words, asking for idle state “X” at the “core”
level gives the hardware a license to go as deep as to idle state “MX” at the
“module” level, but there is no guarantee that this is going to happen (the core
asking for idle state “X” may just end up in that state by itself instead).
Then, the target residency of the
struct cpuidle_state object representing
idle state “X” must reflect the minimum time to spend in idle state “MX” of
the module (including the time needed to enter it), because that is the minimum
time the CPU needs to be idle to save any energy in case the hardware enters
that state. Analogously, the exit latency parameter of that object must cover
the exit time of idle state “MX” of the module (and usually its entry time too),
because that is the maximum delay between a wakeup signal and the time the CPU
will start to execute the first new instruction (assuming that both cores in the
module will always be ready to execute instructions as soon as the module
becomes operational as a whole).
There are processors without direct coordination between different levels of the
hierarchy of units inside them, however. In those cases asking for an idle
state at the “core” level does not automatically affect the “module” level, for
example, in any way and the
CPUIdle driver is responsible for the entire
handling of the hierarchy. Then, the definition of the idle state objects is
entirely up to the driver, but still the physical properties of the idle state
that the processor hardware finally goes into must always follow the parameters
used by the governor for idle state selection (for instance, the actual exit
latency of that idle state must not exceed the exit latency parameter of the
idle state object selected by the governor).
In addition to the target residency and exit latency idle state parameters
discussed above, the objects representing idle states each contain a few other
parameters describing the idle state and a pointer to the function to run in
order to ask the hardware to enter that state. Also, for each
struct cpuidle_state object, there is a corresponding
struct cpuidle_state_usage one containing usage
statistics of the given idle state. That information is exposed by the kernel
For each CPU in the system, there is a
sysfs, where the number
<N> is assigned to the given
CPU at the initialization time. That directory contains a set of subdirectories
state1 and so on, up to the number of idle state
objects defined for the given CPU minus one. Each of these directories
corresponds to one idle state object and the larger the number in its name, the
deeper the (effective) idle state represented by it. Each of them contains
a number of files (attributes) representing the properties of the idle state
object corresponding to it, as follows:
- Total number of times this idle state had been asked for, but the observed idle duration was certainly too short to match its target residency.
- Total number of times this idle state had been asked for, but cerainly a deeper idle state would have been a better match for the observed idle duration.
- Description of the idle state.
- Whether or not this idle state is disabled.
- Exit latency of the idle state in microseconds.
- Name of the idle state.
- Power drawn by hardware in this idle state in milliwatts (if specified, 0 otherwise).
- Target residency of the idle state in microseconds.
- Total time spent in this idle state by the given CPU (as measured by the kernel) in microseconds.
- Total number of times the hardware has been asked by the given CPU to enter this idle state.
name files both contain strings. The difference
between them is that the name is expected to be more concise, while the
description may be longer and it may contain white space or special characters.
The other files listed above contain integer numbers.
disable attribute is the only writeable one. If it contains 1, the
given idle state is disabled for this particular CPU, which means that the
governor will never select it for this particular CPU and the
driver will never ask the hardware to enter it for that CPU as a result.
However, disabling an idle state for one CPU does not prevent it from being
asked for by the other CPUs, so it must be disabled for all of them in order to
never be asked for by any of them. [Note that, due to the way the
governor is implemented, disabling an idle state prevents that governor from
selecting any idle states deeper than the disabled one too.]
disable attribute contains 0, the given idle state is enabled for
this particular CPU, but it still may be disabled for some or all of the other
CPUs in the system at the same time. Writing 1 to it causes the idle state to
be disabled for this particular CPU and writing 0 to it allows the governor to
take it into consideration for the given CPU and the driver to ask for it,
unless that state was disabled globally in the driver (in which case it cannot
be used at all).
power attribute is not defined very well, especially for idle state
objects representing combinations of idle states at different levels of the
hierarchy of units in the processor, and it generally is hard to obtain idle
state power numbers for complex hardware, so
power often contains 0 (not
available) and if it contains a nonzero number, that number may not be very
accurate and it should not be relied on for anything meaningful.
The number in the
time file generally may be greater than the total time
really spent by the given CPU in the given idle state, because it is measured by
the kernel and it may not cover the cases in which the hardware refused to enter
this idle state and entered a shallower one instead of it (or even it did not
enter any idle state at all). The kernel can only measure the time span between
asking the hardware to enter an idle state and the subsequent wakeup of the CPU
and it cannot say what really happened in the meantime at the hardware level.
Moreover, if the idle state object in question represents a combination of idle
states at different levels of the hierarchy of units in the processor,
the kernel can never say how deep the hardware went down the hierarchy in any
particular case. For these reasons, the only reliable way to find out how
much time has been spent by the hardware in different idle states supported by
it is to use idle state residency counters in the hardware, if available.
Power Management Quality of Service for CPUs¶
The power management quality of service (PM QoS) framework in the Linux kernel allows kernel code and user space processes to set constraints on various energy-efficiency features of the kernel to prevent performance from dropping below a required level. The PM QoS constraints can be set globally, in predefined categories referred to as PM QoS classes, or against individual devices.
CPU idle time management can be affected by PM QoS in two ways, through the
global constraint in the
PM_QOS_CPU_DMA_LATENCY class and through the
resume latency constraints for individual CPUs. Kernel code (e.g. device
drivers) can set both of them with the help of special internal interfaces
provided by the PM QoS framework. User space can modify the former by opening
cpu_dma_latency special device file under
/dev/ and writing
a binary value (interpreted as a signed 32-bit integer) to it. In turn, the
resume latency constraint for a CPU can be modified by user space by writing a
string (representing a signed 32-bit integer) to the
power/pm_qos_resume_latency_us file under
sysfs, where the CPU number
<N> is allocated at the system initialization time. Negative values
will be rejected in both cases and, also in both cases, the written integer
number will be interpreted as a requested PM QoS constraint in microseconds.
The requested value is not automatically applied as a new constraint, however, as it may be less restrictive (greater in this particular case) than another constraint previously requested by someone else. For this reason, the PM QoS framework maintains a list of requests that have been made so far in each global class and for each device, aggregates them and applies the effective (minimum in this particular case) value as the new constraint.
In fact, opening the
cpu_dma_latency special device file causes a new
PM QoS request to be created and added to the priority list of requests in the
PM_QOS_CPU_DMA_LATENCY class and the file descriptor coming from the
“open” operation represents that request. If that file descriptor is then
used for writing, the number written to it will be associated with the PM QoS
request represented by it as a new requested constraint value. Next, the
priority list mechanism will be used to determine the new effective value of
the entire list of requests and that effective value will be set as a new
constraint. Thus setting a new requested constraint value will only change the
real constraint if the effective “list” value is affected by it. In particular,
PM_QOS_CPU_DMA_LATENCY class it only affects the real constraint if
it is the minimum of the requested constraints in the list. The process holding
a file descriptor obtained by opening the
cpu_dma_latency special device
file controls the PM QoS request associated with that file descriptor, but it
controls this particular PM QoS request only.
cpu_dma_latency special device file or, more precisely, the
file descriptor obtained while opening it, causes the PM QoS request associated
with that file descriptor to be removed from the
class priority list and destroyed. If that happens, the priority list mechanism
will be used, again, to determine the new effective value for the whole list
and that value will become the new real constraint.
In turn, for each CPU there is only one resume latency PM QoS request
associated with the
power/pm_qos_resume_latency_us file under
sysfs and writing to it causes
this single PM QoS request to be updated regardless of which user space
process does that. In other words, this PM QoS request is shared by the entire
user space, so access to the file associated with it needs to be arbitrated
to avoid confusion. [Arguably, the only legitimate use of this mechanism in
practice is to pin a process to the CPU in question and let it use the
sysfs interface to control the resume latency constraint for it.] It
still only is a request, however. It is a member of a priority list used to
determine the effective value to be set as the resume latency constraint for the
CPU in question every time the list of requests is updated this way or another
(there may be other requests coming from kernel code in that list).
CPU idle time governors are expected to regard the minimum of the global
PM_QOS_CPU_DMA_LATENCY class constraint and the effective
resume latency constraint for the given CPU as the upper limit for the exit
latency of the idle states they can select for that CPU. They should never
select any idle states with exit latency beyond that limit.
Idle States Control Via Kernel Command Line¶
In addition to the
sysfs interface allowing individual idle states to be
disabled for individual CPUs, there are kernel
command line parameters affecting CPU idle time management.
cpuidle.off=1 kernel command line option can be used to disable the
CPU idle time management entirely. It does not prevent the idle loop from
running on idle CPUs, but it prevents the CPU idle time governors and drivers
from being invoked. If it is added to the kernel command line, the idle loop
will ask the hardware to enter idle states on idle CPUs via the CPU architecture
support code that is expected to provide a default mechanism for this purpose.
That default mechanism usually is the least common denominator for all of the
processors implementing the architecture (i.e. CPU instruction set) in question,
however, so it is rather crude and not very energy-efficient. For this reason,
it is not recommended for production use.
cpuidle.governor= kernel command line switch allows the
governor to use to be specified. It has to be appended with a string matching
the name of an available governor (e.g.
cpuidle.governor=menu) and that
governor will be used instead of the default one. It is possible to force
menu governor to be used on the systems that use the
by default this way, for example.
The other kernel command line parameters controlling CPU idle time management described below are only relevant for the x86 architecture and some of them affect Intel processors only.
The x86 architecture support code recognizes three kernel command line
options related to CPU idle time management:
idle=nomwait. The first two of them disable the
intel_idle drivers altogether, which effectively causes the entire
CPUIdle subsystem to be disabled and makes the idle loop invoke the
architecture support code to deal with idle CPUs. How it does that depends on
which of the two parameters is added to the kernel command line. In the
idle=halt case, the architecture support code will use the
instruction of the CPUs (which, as a rule, suspends the execution of the program
and causes the hardware to attempt to enter the shallowest available idle state)
for this purpose, and if
idle=poll is used, idle CPUs will execute a
more or less
lightweight'' sequence of instructions in a tight loop. [Note
that using ``idle=poll is somewhat drastic in many cases, as preventing idle
CPUs from saving almost any energy at all may not be the only effect of it.
For example, on Intel hardware it effectively prevents CPUs from using
P-states (see CPU Performance Scaling) that require any number of CPUs in a package to be
idle, so it very well may hurt single-thread computations performance as well as
energy-efficiency. Thus using it for performance reasons may not be a good idea
idle=nomwait option disables the
intel_idle driver and causes
acpi_idle to be used (as long as all of the information needed by it is
there in the system’s ACPI tables), but it is not allowed to use the
MWAIT instruction of the CPUs to ask the hardware to enter idle states.
In addition to the architecture-level kernel command line options affecting CPU
idle time management, there are parameters affecting individual
drivers that can be passed to them via the kernel command line. Specifically,
<n> is an idle state index also used in the name of the given
state’s directory in
Representation of Idle States), causes the
acpi_idle drivers, respectively, to discard all of the
idle states deeper than idle state
<n>. In that case, they will never ask
for any of those idle states or expose them to the governor. [The behavior of
the two drivers is different for
<n> equal to
intel_idle.max_cstate=0 to the kernel command line disables the
intel_idle driver and allows
acpi_idle to be used, whereas
processor.max_cstate=0 is equivalent to
acpi_idle driver is part of the
processor kernel module that
can be loaded separately and
max_cstate=<n> can be passed to it as a module
parameter when it is loaded.]