x86/idle: Mark "idle=poll" as deprecated
Subsystems (3)
PATCH 1/1 x86/idle: Mark "idle=poll" as deprecated
The "idle=poll" boot parameter is a blunt instrument that forces all CPUs in the system into a continuous "polling" state. While effective at eliminating wake-up latency, this global override is architecturally obsolete and inefficient on modern multicore systems.
It suffers from several significant drawbacks:
1. Lack of Granularity: It prevents "housekeeping" CPUs from entering power-saving states, leading to unnecessary energy waste and thermal throttling that can negatively impact the very latency-sensitive tasks it aims to protect
2. Resource Contention: On SMT systems, a polling sibling thread actively consumes execution resources, potentially degrading the performance of the primary thread on the same physical core.
The Power Management Quality of Service (PM QoS) subsystem now provides a superior, granular alternative.
By writing special value "n/a" to the per-CPU sysfs node /sys/devices/system/cpu/cpuN/power/pm_qos_resume_latency_us, userspace can force a specific CPU to poll without imposing this cost globally. Writing "0" to the same file removes the constraint, allowing the governor to freely select the deepest applicable C-state.
This patch marks "idle=poll" as deprecated. A warning is issued at boot time if the parameter is used, guiding users toward the PM QoS interface.
--- Documentation/admin-guide/kernel-parameters.txt | 9 +++++---- Documentation/timers/no_hz.rst | 9 +++++++++ arch/x86/kernel/process.c | 1 + 3 files changed, 15 insertions(+), 4 deletions(-)
Documentation/admin-guide/kernel-parameters.txt +5 -4
@@ -2242,10 +2242,11 @@ Kernel parameters idle= [X86,EARLY] Format: idle=poll, idle=halt, idle=nomwait - idle=poll: Don't do power saving in the idle loop - using HLT, but poll for rescheduling event. This will - make the CPUs eat a lot more power, but may be useful - to get slightly better performance in multiprocessor + idle=poll: [Deprecated - use PM QoS] + Don't do power saving in the idle loop using HLT, but + poll for rescheduling event. This will make the CPUs + eat a lot more power, but may be useful to get + slightly better performance in multiprocessor benchmarks. It also makes some profiling using performance counters more accurate. Please note that on systems with MONITOR/MWAIT support (like Intel
Documentation/timers/no_hz.rst +9 -0
@@ -234,6 +234,15 @@ Known Issues a. Use PMQOS from userspace to inform the kernel of your latency requirements (preferred). + This interface offers a superior and more flexible alternative to + global boot parameters such as "idle=poll", as it can be adjusted + at runtime with per-CPU granularity. + + To force a specific CPU (where N is the logical CPU number) to poll + on idle, one can set the latency requirement to 0 microseconds: + + echo "n/a" > /sys/devices/system/cpu/cpuN/power/pm_qos_resume_latency_us + b. On x86 systems, use the "idle=mwait" boot parameter. c. On x86 systems, use the "intel_idle.max_cstate=" to limit
arch/x86/kernel/process.c +1 -0
@@ -1000,6 +1000,7 @@ static int __init idle_setup(char *str)if (!strcmp (str,"poll" )) {pr_info ("using polling idle threads\n" ); +pr_warn ("idle=poll is deprecated. Use the PM QoS interface instead via /sys/devices/system/cpu/cpuN/power/\n" ); boot_option_idle_override =IDLE_POLL ;cpu_idle_poll_ctrl (true); }else if (!strcmp (str,"halt" )) {
Signature · 1 line
2.51.0
Borislav Petkov
maintainerQuoted 4 lines · Aaron Tomlin
The "idle=poll" boot parameter is a blunt instrument that forces all CPUs in the system into a continuous "polling" state. While effective at eliminating wake-up latency, this global override is architecturally obsolete and inefficient on modern multicore systems.
Nope.
It is still very useful when you want to stop the machine from entering sleep states.
Quoted 3 lines
By writing special value "n/a" to the per-CPU sysfs node
/sys/devices/system/cpu/cpuN/power/pm_qos_resume_latency_us, userspace can force a specific CPU to poll without imposing this cost globally.
You can't do that on the kernel command line and thus prevent entering sleep states from the get-go.
So no, it ain't broke so no need to fix it.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteQuoted 11 lines · Borislav Petkov
Quoted 4 lines · Aaron Tomlin
The "idle=poll" boot parameter is a blunt instrument that forces all CPUs in the system into a continuous "polling" state. While effective at eliminating wake-up latency, this global override is architecturally obsolete and inefficient on modern multicore systems.
Nope.
It is still very useful when you want to stop the machine from entering sleep states.
Quoted 6 lines
Quoted 3 lines
By writing special value "n/a" to the per-CPU sysfs node
/sys/devices/system/cpu/cpuN/power/pm_qos_resume_latency_us, userspace can force a specific CPU to poll without imposing this cost globally.You can't do that on the kernel command line and thus prevent entering sleep states from the get-go.
Hi Boris,
Thank you for your feedback.
Understood.
I will investigate the possibility of implementing a boot-time parameter for the PM QoS subsystem to bridge this gap. This would provide the same immediate enforcement as "idle=poll" while retaining the runtime flexibility of the QoS framework.
Kind rgards,
Signature · 1 line
Aaron Tomlin
Borislav Petkov
maintainerQuoted 4 lines · Aaron Tomlin
I will investigate the possibility of implementing a boot-time parameter for the PM QoS subsystem to bridge this gap. This would provide the same immediate enforcement as "idle=poll" while retaining the runtime flexibility of the QoS framework.
No, if what you claim is a valid and actual use case, then all it needs is documentation. Everything else is unnecessary.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette