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, 191 insertions, 0 deletions

diff --git a/Documentation/driver-api/pwm.rst b/Documentation/driver-api/pwm.rst
new file mode 100644
index 000000000..0d27a40f5
--- /dev/null
+++ b/Documentation/driver-api/pwm.rst
@@ -0,0 +1,191 @@
+======================================
+Pulse Width Modulation (PWM) interface
+======================================
+
+This provides an overview about the Linux PWM interface
+
+PWMs are commonly used for controlling LEDs, fans or vibrators in
+cell phones. PWMs with a fixed purpose have no need implementing
+the Linux PWM API (although they could). However, PWMs are often
+found as discrete devices on SoCs which have no fixed purpose. It's
+up to the board designer to connect them to LEDs or fans. To provide
+this kind of flexibility the generic PWM API exists.
+
+Identifying PWMs
+----------------
+
+Users of the legacy PWM API use unique IDs to refer to PWM devices.
+
+Instead of referring to a PWM device via its unique ID, board setup code
+should instead register a static mapping that can be used to match PWM
+consumers to providers, as given in the following example::
+
+	static struct pwm_lookup board_pwm_lookup[] = {
+		PWM_LOOKUP("tegra-pwm", 0, "pwm-backlight", NULL,
+			   50000, PWM_POLARITY_NORMAL),
+	};
+
+	static void __init board_init(void)
+	{
+		...
+		pwm_add_table(board_pwm_lookup, ARRAY_SIZE(board_pwm_lookup));
+		...
+	}
+
+Using PWMs
+----------
+
+Consumers use the pwm_get() function and pass to it the consumer device or a
+consumer name. pwm_put() is used to free the PWM device. Managed variants of
+the getter, devm_pwm_get() and devm_fwnode_pwm_get(), also exist.
+
+After being requested, a PWM has to be configured using::
+
+	int pwm_apply_might_sleep(struct pwm_device *pwm, struct pwm_state *state);
+
+This API controls both the PWM period/duty_cycle config and the
+enable/disable state.
+
+PWM devices can be used from atomic context, if the PWM does not sleep. You
+can check if this the case with::
+
+        bool pwm_might_sleep(struct pwm_device *pwm);
+
+If false, the PWM can also be configured from atomic context with::
+
+	int pwm_apply_atomic(struct pwm_device *pwm, struct pwm_state *state);
+
+As a consumer, don't rely on the output's state for a disabled PWM. If it's
+easily possible, drivers are supposed to emit the inactive state, but some
+drivers cannot. If you rely on getting the inactive state, use .duty_cycle=0,
+.enabled=true.
+
+There is also a usage_power setting: If set, the PWM driver is only required to
+maintain the power output but has more freedom regarding signal form.
+If supported by the driver, the signal can be optimized, for example to improve
+EMI by phase shifting the individual channels of a chip.
+
+The pwm_config(), pwm_enable() and pwm_disable() functions are just wrappers
+around pwm_apply_might_sleep() and should not be used if the user wants to change
+several parameter at once. For example, if you see pwm_config() and
+pwm_{enable,disable}() calls in the same function, this probably means you
+should switch to pwm_apply_might_sleep().
+
+The PWM user API also allows one to query the PWM state that was passed to the
+last invocation of pwm_apply_might_sleep() using pwm_get_state(). Note this is
+different to what the driver has actually implemented if the request cannot be
+satisfied exactly with the hardware in use. There is currently no way for
+consumers to get the actually implemented settings.
+
+In addition to the PWM state, the PWM API also exposes PWM arguments, which
+are the reference PWM config one should use on this PWM.
+PWM arguments are usually platform-specific and allows the PWM user to only
+care about dutycycle relatively to the full period (like, duty = 50% of the
+period). struct pwm_args contains 2 fields (period and polarity) and should
+be used to set the initial PWM config (usually done in the probe function
+of the PWM user). PWM arguments are retrieved with pwm_get_args().
+
+All consumers should really be reconfiguring the PWM upon resume as
+appropriate. This is the only way to ensure that everything is resumed in
+the proper order.
+
+Using PWMs with the sysfs interface
+-----------------------------------
+
+If CONFIG_SYSFS is enabled in your kernel configuration a simple sysfs
+interface is provided to use the PWMs from userspace. It is exposed at
+/sys/class/pwm/. Each probed PWM controller/chip will be exported as
+pwmchipN, where N is the base of the PWM chip. Inside the directory you
+will find:
+
+  npwm
+    The number of PWM channels this chip supports (read-only).
+
+  export
+    Exports a PWM channel for use with sysfs (write-only).
+
+  unexport
+   Unexports a PWM channel from sysfs (write-only).
+
+The PWM channels are numbered using a per-chip index from 0 to npwm-1.
+
+When a PWM channel is exported a pwmX directory will be created in the
+pwmchipN directory it is associated with, where X is the number of the
+channel that was exported. The following properties will then be available:
+
+  period
+    The total period of the PWM signal (read/write).
+    Value is in nanoseconds and is the sum of the active and inactive
+    time of the PWM.
+
+  duty_cycle
+    The active time of the PWM signal (read/write).
+    Value is in nanoseconds and must be less than or equal to the period.
+
+  polarity
+    Changes the polarity of the PWM signal (read/write).
+    Writes to this property only work if the PWM chip supports changing
+    the polarity.
+    Value is the string "normal" or "inversed".
+
+  enable
+    Enable/disable the PWM signal (read/write).
+
+	- 0 - disabled
+	- 1 - enabled
+
+Implementing a PWM driver
+-------------------------
+
+Currently there are two ways to implement pwm drivers. Traditionally
+there only has been the barebone API meaning that each driver has
+to implement the pwm_*() functions itself. This means that it's impossible
+to have multiple PWM drivers in the system. For this reason it's mandatory
+for new drivers to use the generic PWM framework.
+
+A new PWM controller/chip can be allocated using pwmchip_alloc(), then
+registered using pwmchip_add() and removed again with pwmchip_remove(). To undo
+pwmchip_alloc() use pwmchip_put(). pwmchip_add() takes a filled in struct
+pwm_chip as argument which provides a description of the PWM chip, the number
+of PWM devices provided by the chip and the chip-specific implementation of the
+supported PWM operations to the framework.
+
+When implementing polarity support in a PWM driver, make sure to respect the
+signal conventions in the PWM framework. By definition, normal polarity
+characterizes a signal starts high for the duration of the duty cycle and
+goes low for the remainder of the period. Conversely, a signal with inversed
+polarity starts low for the duration of the duty cycle and goes high for the
+remainder of the period.
+
+Drivers are encouraged to implement ->apply() instead of the legacy
+->enable(), ->disable() and ->config() methods. Doing that should provide
+atomicity in the PWM config workflow, which is required when the PWM controls
+a critical device (like a regulator).
+
+The implementation of ->get_state() (a method used to retrieve initial PWM
+state) is also encouraged for the same reason: letting the PWM user know
+about the current PWM state would allow him to avoid glitches.
+
+Drivers should not implement any power management. In other words,
+consumers should implement it as described in the "Using PWMs" section.
+
+Locking
+-------
+
+The PWM core list manipulations are protected by a mutex, so pwm_get()
+and pwm_put() may not be called from an atomic context.
+Most functions in the PWM consumer API might sleep and so must not be called
+from atomic context. The notable exception is pwm_apply_atomic() which has the
+same semantics as pwm_apply_might_sleep() but can be called from atomic context.
+(The price for that is that it doesn't work for all PWM devices, use
+pwm_might_sleep() to check if a given PWM supports atomic operation.
+
+Locking in the PWM core ensures that callbacks related to a single chip are
+serialized.
+
+Helpers
+-------
+
+Currently a PWM can only be configured with period_ns and duty_ns. For several
+use cases freq_hz and duty_percent might be better. Instead of calculating
+this in your driver please consider adding appropriate helpers to the framework.