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, 150 insertions, 0 deletions

diff --git a/Documentation/devicetree/overlay-notes.rst b/Documentation/devicetree/overlay-notes.rst
new file mode 100644
index 000000000..ba401ef85
--- /dev/null
+++ b/Documentation/devicetree/overlay-notes.rst
@@ -0,0 +1,150 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+========================
+Devicetree Overlay Notes
+========================
+
+This document describes the implementation of the in-kernel
+device tree overlay functionality residing in drivers/of/overlay.c and is a
+companion document to Documentation/devicetree/dynamic-resolution-notes.rst[1]
+
+How overlays work
+-----------------
+
+A Devicetree's overlay purpose is to modify the kernel's live tree, and
+have the modification affecting the state of the kernel in a way that
+is reflecting the changes.
+Since the kernel mainly deals with devices, any new device node that results
+in an active device should have it created while if the device node is either
+disabled or removed all together, the affected device should be deregistered.
+
+Let's take an example where we have a foo board with the following base tree::
+
+    ---- foo.dts ---------------------------------------------------------------
+	/* FOO platform */
+	/dts-v1/;
+	/ {
+		compatible = "corp,foo";
+
+		/* shared resources */
+		res: res {
+		};
+
+		/* On chip peripherals */
+		ocp: ocp {
+			/* peripherals that are always instantiated */
+			peripheral1 { ... };
+		};
+	};
+    ---- foo.dts ---------------------------------------------------------------
+
+The overlay bar.dtso,
+::
+
+    ---- bar.dtso - overlay target location by label ---------------------------
+	/dts-v1/;
+	/plugin/;
+	&ocp {
+		/* bar peripheral */
+		bar {
+			compatible = "corp,bar";
+			... /* various properties and child nodes */
+		};
+	};
+    ---- bar.dtso --------------------------------------------------------------
+
+when loaded (and resolved as described in [1]) should result in foo+bar.dts::
+
+    ---- foo+bar.dts -----------------------------------------------------------
+	/* FOO platform + bar peripheral */
+	/ {
+		compatible = "corp,foo";
+
+		/* shared resources */
+		res: res {
+		};
+
+		/* On chip peripherals */
+		ocp: ocp {
+			/* peripherals that are always instantiated */
+			peripheral1 { ... };
+
+			/* bar peripheral */
+			bar {
+				compatible = "corp,bar";
+				... /* various properties and child nodes */
+			};
+		};
+	};
+    ---- foo+bar.dts -----------------------------------------------------------
+
+As a result of the overlay, a new device node (bar) has been created
+so a bar platform device will be registered and if a matching device driver
+is loaded the device will be created as expected.
+
+If the base DT was not compiled with the -@ option then the "&ocp" label
+will not be available to resolve the overlay node(s) to the proper location
+in the base DT. In this case, the target path can be provided. The target
+location by label syntax is preferred because the overlay can be applied to
+any base DT containing the label, no matter where the label occurs in the DT.
+
+The above bar.dtso example modified to use target path syntax is::
+
+    ---- bar.dtso - overlay target location by explicit path -------------------
+	/dts-v1/;
+	/plugin/;
+	&{/ocp} {
+		/* bar peripheral */
+		bar {
+			compatible = "corp,bar";
+			... /* various properties and child nodes */
+		}
+	};
+    ---- bar.dtso --------------------------------------------------------------
+
+
+Overlay in-kernel API
+--------------------------------
+
+The API is quite easy to use.
+
+1) Call of_overlay_fdt_apply() to create and apply an overlay changeset. The
+   return value is an error or a cookie identifying this overlay.
+
+2) Call of_overlay_remove() to remove and clean up the overlay changeset
+   previously created via the call to of_overlay_fdt_apply(). Removal of an
+   overlay changeset that is stacked by another will not be permitted.
+
+Finally, if you need to remove all overlays in one-go, just call
+of_overlay_remove_all() which will remove every single one in the correct
+order.
+
+There is the option to register notifiers that get called on
+overlay operations. See of_overlay_notifier_register/unregister and
+enum of_overlay_notify_action for details.
+
+A notifier callback for OF_OVERLAY_PRE_APPLY, OF_OVERLAY_POST_APPLY, or
+OF_OVERLAY_PRE_REMOVE may store pointers to a device tree node in the overlay
+or its content but these pointers must not persist past the notifier callback
+for OF_OVERLAY_POST_REMOVE.  The memory containing the overlay will be
+kfree()ed after OF_OVERLAY_POST_REMOVE notifiers are called.  Note that the
+memory will be kfree()ed even if the notifier for OF_OVERLAY_POST_REMOVE
+returns an error.
+
+The changeset notifiers in drivers/of/dynamic.c are a second type of notifier
+that could be triggered by applying or removing an overlay.  These notifiers
+are not allowed to store pointers to a device tree node in the overlay
+or its content.  The overlay code does not protect against such pointers
+remaining active when the memory containing the overlay is freed as a result
+of removing the overlay.
+
+Any other code that retains a pointer to the overlay nodes or data is
+considered to be a bug because after removing the overlay the pointer
+will refer to freed memory.
+
+Users of overlays must be especially aware of the overall operations that
+occur on the system to ensure that other kernel code does not retain any
+pointers to the overlay nodes or data.  Any example of an inadvertent use
+of such pointers is if a driver or subsystem module is loaded after an
+overlay has been applied, and the driver or subsystem scans the entire
+devicetree or a large portion of it, including the overlay nodes.