sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling buggrafted

alb_break_llc() decides whether to break LLC preference during active
load balance. It does so by testing that every runnable fair task on the
source rq prefers its LLC:

	env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable

But the two counters cover different sets. nr_pref_llc_running is updated
in account_llc_enqueue()/account_llc_dequeue(), next to cfs_rq->nr_queued,
so it follows queued tasks. h_nr_runnable is updated in set_delayed()/
clear_delayed() and drops delay-dequeued tasks.

So under DELAY_DEQUEUE, a preferring task that goes to sleep stays counted
in nr_pref_llc_running while h_nr_runnable falls. The equality then breaks,
alb_break_llc() returns false, and active balance is free to pull a task
off its preferred LLC. Active balance only moves runnable tasks, and this
is the only LLC check it consults: once the stopper runs, LBF_ACTIVE_LB
skips the per-task test in can_migrate_task(). The runnable set is the one
we want.

Fix it on the counter side. A task should be counted in
nr_pref_llc_running exactly while it is both queued on its preferred LLC
(pref_llc_queued) and runnable (!sched_delayed). Define that membership
once in task_pref_llc_runnable(), and adjust the counter only through
pref_llc_running_inc()/pref_llc_running_dec() from the four sites that
change either input: account_llc_enqueue(), account_llc_dequeue(),
set_delayed() and clear_delayed(). Gating every update on the same
predicate keeps the delay, wake and dequeue paths from double-counting
or underflowing; see the comments at those sites for the ordering.

nr_llc_running and sd->llc_counts are not touched and stay on queued
semantics.

Fixes: 714059f79ff0 ("sched/cache: Handle moving single tasks to/from their preferred LLC")
Closes: https://lore.kernel.org/lkml/20260827135000.735138-1-zhanxusheng@xiaomi.com/
Reported-by: Zhan Xusheng <zhanxusheng@xiaomi.com>
Suggested-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kayra Cizmeci <kayracizmeci@gmail.com>
Cc: <stable@kernel.org> # v7.2.x
Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com

1 files changed, 107 insertions, 0 deletions

diff --git a/Documentation/driver-api/thermal/cpu-cooling-api.rst b/Documentation/driver-api/thermal/cpu-cooling-api.rst
new file mode 100644
index 000000000..645d914c4
--- /dev/null
+++ b/Documentation/driver-api/thermal/cpu-cooling-api.rst
@@ -0,0 +1,107 @@
+=======================
+CPU cooling APIs How To
+=======================
+
+Written by Amit Daniel Kachhap <amit.kachhap@linaro.org>
+
+Updated: 6 Jan 2015
+
+Copyright (c)  2012 Samsung Electronics Co., Ltd(http://www.samsung.com)
+
+0. Introduction
+===============
+
+The generic cpu cooling(freq clipping) provides registration/unregistration APIs
+to the caller. The binding of the cooling devices to the trip point is left for
+the user. The registration APIs returns the cooling device pointer.
+
+1. cpu cooling APIs
+===================
+
+1.1 cpufreq registration/unregistration APIs
+--------------------------------------------
+
+    ::
+
+	struct thermal_cooling_device
+	*cpufreq_cooling_register(struct cpumask *clip_cpus)
+
+    This interface function registers the cpufreq cooling device with the name
+    "thermal-cpufreq-%x". This api can support multiple instances of cpufreq
+    cooling devices.
+
+   clip_cpus:
+	cpumask of cpus where the frequency constraints will happen.
+
+    ::
+
+	struct thermal_cooling_device
+	*of_cpufreq_cooling_register(struct cpufreq_policy *policy)
+
+    This interface function registers the cpufreq cooling device with
+    the name "thermal-cpufreq-%x" linking it with a device tree node, in
+    order to bind it via the thermal DT code. This api can support multiple
+    instances of cpufreq cooling devices.
+
+    policy:
+	CPUFreq policy.
+
+
+    ::
+
+	void cpufreq_cooling_unregister(struct thermal_cooling_device *cdev)
+
+    This interface function unregisters the "thermal-cpufreq-%x" cooling device.
+
+    cdev: Cooling device pointer which has to be unregistered.
+
+2. Power models
+===============
+
+The power API registration functions provide a simple power model for
+CPUs.  The current power is calculated as dynamic power (static power isn't
+supported currently).  This power model requires that the operating-points of
+the CPUs are registered using the kernel's opp library and the
+`cpufreq_frequency_table` is assigned to the `struct device` of the
+cpu.  If you are using CONFIG_CPUFREQ_DT then the
+`cpufreq_frequency_table` should already be assigned to the cpu
+device.
+
+The dynamic power consumption of a processor depends on many factors.
+For a given processor implementation the primary factors are:
+
+- The time the processor spends running, consuming dynamic power, as
+  compared to the time in idle states where dynamic consumption is
+  negligible.  Herein we refer to this as 'utilisation'.
+- The voltage and frequency levels as a result of DVFS.  The DVFS
+  level is a dominant factor governing power consumption.
+- In running time the 'execution' behaviour (instruction types, memory
+  access patterns and so forth) causes, in most cases, a second order
+  variation.  In pathological cases this variation can be significant,
+  but typically it is of a much lesser impact than the factors above.
+
+A high level dynamic power consumption model may then be represented as::
+
+	Pdyn = f(run) * Voltage^2 * Frequency * Utilisation
+
+f(run) here represents the described execution behaviour and its
+result has a units of Watts/Hz/Volt^2 (this often expressed in
+mW/MHz/uVolt^2)
+
+The detailed behaviour for f(run) could be modelled on-line.  However,
+in practice, such an on-line model has dependencies on a number of
+implementation specific processor support and characterisation
+factors.  Therefore, in initial implementation that contribution is
+represented as a constant coefficient.  This is a simplification
+consistent with the relative contribution to overall power variation.
+
+In this simplified representation our model becomes::
+
+	Pdyn = Capacitance * Voltage^2 * Frequency * Utilisation
+
+Where `capacitance` is a constant that represents an indicative
+running time dynamic power coefficient in fundamental units of
+mW/MHz/uVolt^2.  Typical values for mobile CPUs might lie in range
+from 100 to 500.  For reference, the approximate values for the SoC in
+ARM's Juno Development Platform are 530 for the Cortex-A57 cluster and
+140 for the Cortex-A53 cluster.