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

7 files changed, 156 insertions, 0 deletions

diff --git a/scripts/dummy-tools/dummy-plugin-dir/include/plugin-version.h b/scripts/dummy-tools/dummy-plugin-dir/include/plugin-version.h
new file mode 100644
index 000000000..e69de29bb
--- /dev/null
+++ b/scripts/dummy-tools/dummy-plugin-dir/include/plugin-version.h
diff --git a/scripts/dummy-tools/gcc b/scripts/dummy-tools/gcc
new file mode 100755
index 000000000..e6c41427c
--- /dev/null
+++ b/scripts/dummy-tools/gcc
@@ -0,0 +1,116 @@
+#!/bin/sh
+# SPDX-License-Identifier: GPL-2.0-only
+#
+# Staring v4.18, Kconfig evaluates compiler capabilities, and hides CONFIG
+# options your compiler does not support. This works well if you configure and
+# build the kernel on the same host machine.
+#
+# It is inconvenient if you prepare the .config that is carried to a different
+# build environment (typically this happens when you package the kernel for
+# distros) because using a different compiler potentially produces different
+# CONFIG options than the real build environment. So, you probably want to make
+# as many options visible as possible. In other words, you need to create a
+# super-set of CONFIG options that cover any build environment. If some of the
+# CONFIG options turned out to be unsupported on the build machine, they are
+# automatically disabled by the nature of Kconfig.
+#
+# However, it is not feasible to get a full-featured compiler for every arch.
+# Hence these dummy toolchains to make all compiler tests pass.
+#
+# Usage:
+#
+# From the top directory of the source tree, run
+#
+#   $ make CROSS_COMPILE=scripts/dummy-tools/ oldconfig
+#
+# Most of compiler features are tested by cc-option, which simply checks the
+# exit code of $(CC). This script does nothing and just exits with 0 in most
+# cases. So, $(cc-option, ...) is evaluated as 'y'.
+#
+# This scripts caters to more checks; handle --version and pre-process __GNUC__
+# etc. to pretend to be GCC, and also do right things to satisfy some scripts.
+
+# Check if the first parameter appears in the rest. Succeeds if found.
+# This helper is useful if a particular option was passed to this script.
+# Typically used like this:
+#   arg_contain <word-you-are-searching-for> "$@"
+arg_contain ()
+{
+	search="$1"
+	shift
+
+	while [ $# -gt 0 ]
+	do
+		if [ "$search" = "$1" ]; then
+			return 0
+		fi
+		shift
+	done
+
+	return 1
+}
+
+# To set CONFIG_CC_IS_GCC=y
+if arg_contain --version "$@"; then
+	echo "gcc (scripts/dummy-tools/gcc)"
+	exit 0
+fi
+
+if arg_contain -E "$@"; then
+	# For scripts/cc-version.sh; This emulates GCC 20.0.0
+	if arg_contain - "$@"; then
+		sed -n '/^GCC/{s/__GNUC__/20/; s/__GNUC_MINOR__/0/; s/__GNUC_PATCHLEVEL__/0/; p;}; s/__LONG_DOUBLE_128__/1/ p'
+		exit 0
+	else
+		echo "no input files" >&2
+		exit 1
+	fi
+fi
+
+# To set CONFIG_AS_IS_GNU
+if arg_contain -Wa,--version "$@"; then
+	echo "GNU assembler (scripts/dummy-tools) 2.50"
+	exit 0
+fi
+
+if arg_contain -S "$@"; then
+	# For scripts/gcc-x86-*-has-stack-protector.sh
+	if arg_contain -fstack-protector "$@"; then
+		if arg_contain -mstack-protector-guard-reg=fs "$@"; then
+			echo "%fs"
+		else
+			echo "%gs"
+		fi
+		exit 0
+	fi
+
+	# For arch/powerpc/tools/gcc-check-mprofile-kernel.sh
+	if arg_contain -m64 "$@" && arg_contain -mprofile-kernel "$@"; then
+		if ! test -t 0 && ! grep -q notrace; then
+			echo "_mcount"
+		fi
+		exit 0
+	fi
+
+	# For arch/powerpc/tools/gcc-check-fpatchable-function-entry.sh
+	if arg_contain -m64 "$@" && arg_contain -fpatchable-function-entry=2 "$@"; then
+		echo "func:"
+		echo ".section __patchable_function_entries"
+		echo ".localentry"
+		echo "  nop"
+		echo "  nop"
+		exit 0
+	fi
+fi
+
+# To set GCC_PLUGINS
+if arg_contain -print-file-name=plugin "$@"; then
+	# Use $0 to find the in-tree dummy directory
+	echo "$(dirname "$(readlink -f "$0")")/dummy-plugin-dir"
+	exit 0
+fi
+
+# inverted return value
+if arg_contain -D__SIZEOF_INT128__=0 "$@"; then
+	exit 1
+fi
diff --git a/scripts/dummy-tools/ld b/scripts/dummy-tools/ld
new file mode 100755
index 000000000..f68233050
--- /dev/null
+++ b/scripts/dummy-tools/ld
@@ -0,0 +1,30 @@
+#!/bin/sh
+# SPDX-License-Identifier: GPL-2.0-only
+
+# Dummy script that always succeeds.
+
+# Check if the first parameter appears in the rest. Succeeds if found.
+# This helper is useful if a particular option was passed to this script.
+# Typically used like this:
+#   arg_contain <word-you-are-searching-for> "$@"
+arg_contain ()
+{
+	search="$1"
+	shift
+
+	while [ $# -gt 0 ]
+	do
+		if [ "$search" = "$1" ]; then
+			return 0
+		fi
+		shift
+	done
+
+	return 1
+}
+
+if arg_contain --version "$@" || arg_contain -v "$@"; then
+	progname=$(basename $0)
+	echo "GNU $progname (scripts/dummy-tools/$progname) 2.50"
+	exit 0
+fi
diff --git a/scripts/dummy-tools/nm b/scripts/dummy-tools/nm
new file mode 120000
index 000000000..c0648b38d
--- /dev/null
+++ b/scripts/dummy-tools/nm
@@ -0,0 +1 @@
+ld
\ No newline at end of file
diff --git a/scripts/dummy-tools/objcopy b/scripts/dummy-tools/objcopy
new file mode 120000
index 000000000..c0648b38d
--- /dev/null
+++ b/scripts/dummy-tools/objcopy
@@ -0,0 +1 @@
+ld
\ No newline at end of file
diff --git a/scripts/dummy-tools/pahole b/scripts/dummy-tools/pahole
new file mode 100755
index 000000000..53501a36f
--- /dev/null
+++ b/scripts/dummy-tools/pahole
@@ -0,0 +1,4 @@
+#!/bin/sh
+# SPDX-License-Identifier: GPL-2.0-only
+
+echo v99.99
diff --git a/scripts/dummy-tools/python3 b/scripts/dummy-tools/python3
new file mode 100755
index 000000000..24c558486
--- /dev/null
+++ b/scripts/dummy-tools/python3
@@ -0,0 +1,4 @@
+#!/bin/sh
+# SPDX-License-Identifier: GPL-2.0-only
+
+true