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, 137 insertions, 0 deletions

diff --git a/Documentation/input/devices/ntrig.rst b/Documentation/input/devices/ntrig.rst
new file mode 100644
index 000000000..1559f5349
--- /dev/null
+++ b/Documentation/input/devices/ntrig.rst
@@ -0,0 +1,137 @@
+.. include:: <isonum.txt>
+
+=========================
+N-Trig touchscreen Driver
+=========================
+
+:Copyright: |copy| 2008-2010 Rafi Rubin <rafi@seas.upenn.edu>
+:Copyright: |copy| 2009-2010 Stephane Chatty
+
+This driver provides support for N-Trig pen and multi-touch sensors.  Single
+and multi-touch events are translated to the appropriate protocols for
+the hid and input systems.  Pen events are sufficiently hid compliant and
+are left to the hid core.  The driver also provides additional filtering
+and utility functions accessible with sysfs and module parameters.
+
+This driver has been reported to work properly with multiple N-Trig devices
+attached.
+
+
+Parameters
+----------
+
+Note: values set at load time are global and will apply to all applicable
+devices.  Adjusting parameters with sysfs will override the load time values,
+but only for that one device.
+
+The following parameters are used to configure filters to reduce noise:
+
++-----------------------+-----------------------------------------------------+
+|activate_slack		|number of fingers to ignore before processing events |
++-----------------------+-----------------------------------------------------+
+|activation_height,	|size threshold to activate immediately		      |
+|activation_width	|						      |
++-----------------------+-----------------------------------------------------+
+|min_height,		|size threshold below which fingers are ignored       |
+|min_width		|both to decide activation and during activity	      |
++-----------------------+-----------------------------------------------------+
+|deactivate_slack	|the number of "no contact" frames to ignore before   |
+|			|propagating the end of activity events		      |
++-----------------------+-----------------------------------------------------+
+
+When the last finger is removed from the device, it sends a number of empty
+frames.  By holding off on deactivation for a few frames we can tolerate false
+erroneous disconnects, where the sensor may mistakenly not detect a finger that
+is still present.  Thus deactivate_slack addresses problems where a users might
+see breaks in lines during drawing, or drop an object during a long drag.
+
+
+Additional sysfs items
+----------------------
+
+These nodes just provide easy access to the ranges reported by the device.
+
++-----------------------+-----------------------------------------------------+
+|sensor_logical_height, | the range for positions reported during activity    |
+|sensor_logical_width   |                                                     |
++-----------------------+-----------------------------------------------------+
+|sensor_physical_height,| internal ranges not used for normal events but      |
+|sensor_physical_width  | useful for tuning                                   |
++-----------------------+-----------------------------------------------------+
+
+All N-Trig devices with product id of 1 report events in the ranges of
+
+* X: 0-9600
+* Y: 0-7200
+
+However not all of these devices have the same physical dimensions.  Most
+seem to be 12" sensors (Dell Latitude XT and XT2 and the HP TX2), and
+at least one model (Dell Studio 17) has a 17" sensor.  The ratio of physical
+to logical sizes is used to adjust the size based filter parameters.
+
+
+Filtering
+---------
+
+With the release of the early multi-touch firmwares it became increasingly
+obvious that these sensors were prone to erroneous events.  Users reported
+seeing both inappropriately dropped contact and ghosts, contacts reported
+where no finger was actually touching the screen.
+
+Deactivation slack helps prevent dropped contact for single touch use, but does
+not address the problem of dropping one of more contacts while other contacts
+are still active.  Drops in the multi-touch context require additional
+processing and should be handled in tandem with tacking.
+
+As observed ghost contacts are similar to actual use of the sensor, but they
+seem to have different profiles.  Ghost activity typically shows up as small
+short lived touches.  As such, I assume that the longer the continuous stream
+of events the more likely those events are from a real contact, and that the
+larger the size of each contact the more likely it is real.  Balancing the
+goals of preventing ghosts and accepting real events quickly (to minimize
+user observable latency), the filter accumulates confidence for incoming
+events until it hits thresholds and begins propagating.  In the interest in
+minimizing stored state as well as the cost of operations to make a decision,
+I've kept that decision simple.
+
+Time is measured in terms of the number of fingers reported, not frames since
+the probability of multiple simultaneous ghosts is expected to drop off
+dramatically with increasing numbers.  Rather than accumulate weight as a
+function of size, I just use it as a binary threshold.  A sufficiently large
+contact immediately overrides the waiting period and leads to activation.
+
+Setting the activation size thresholds to large values will result in deciding
+primarily on activation slack.  If you see longer lived ghosts, turning up the
+activation slack while reducing the size thresholds may suffice to eliminate
+the ghosts while keeping the screen quite responsive to firm taps.
+
+Contacts continue to be filtered with min_height and min_width even after
+the initial activation filter is satisfied.  The intent is to provide
+a mechanism for filtering out ghosts in the form of an extra finger while
+you actually are using the screen.  In practice this sort of ghost has
+been far less problematic or relatively rare and I've left the defaults
+set to 0 for both parameters, effectively turning off that filter.
+
+I don't know what the optimal values are for these filters.  If the defaults
+don't work for you, please play with the parameters.  If you do find other
+values more comfortable, I would appreciate feedback.
+
+The calibration of these devices does drift over time.  If ghosts or contact
+dropping worsen and interfere with the normal usage of your device, try
+recalibrating it.
+
+
+Calibration
+-----------
+
+The N-Trig windows tools provide calibration and testing routines.  Also an
+unofficial unsupported set of user space tools including a calibrator is
+available at:
+http://code.launchpad.net/~rafi-seas/+junk/ntrig_calib
+
+
+Tracking
+--------
+
+As of yet, all tested N-Trig firmwares do not track fingers.  When multiple
+contacts are active they seem to be sorted primarily by Y position.