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, 102 insertions, 0 deletions

diff --git a/include/xen/interface/features.h b/include/xen/interface/features.h
new file mode 100644
index 000000000..53f760378
--- /dev/null
+++ b/include/xen/interface/features.h
@@ -0,0 +1,102 @@
+/* SPDX-License-Identifier: MIT */
+/******************************************************************************
+ * features.h
+ *
+ * Feature flags, reported by XENVER_get_features.
+ *
+ * Copyright (c) 2006, Keir Fraser <keir@xensource.com>
+ */
+
+#ifndef __XEN_PUBLIC_FEATURES_H__
+#define __XEN_PUBLIC_FEATURES_H__
+
+/*
+ * If set, the guest does not need to write-protect its pagetables, and can
+ * update them via direct writes.
+ */
+#define XENFEAT_writable_page_tables       0
+
+/*
+ * If set, the guest does not need to write-protect its segment descriptor
+ * tables, and can update them via direct writes.
+ */
+#define XENFEAT_writable_descriptor_tables 1
+
+/*
+ * If set, translation between the guest's 'pseudo-physical' address space
+ * and the host's machine address space are handled by the hypervisor. In this
+ * mode the guest does not need to perform phys-to/from-machine translations
+ * when performing page table operations.
+ */
+#define XENFEAT_auto_translated_physmap    2
+
+/* If set, the guest is running in supervisor mode (e.g., x86 ring 0). */
+#define XENFEAT_supervisor_mode_kernel     3
+
+/*
+ * If set, the guest does not need to allocate x86 PAE page directories
+ * below 4GB. This flag is usually implied by auto_translated_physmap.
+ */
+#define XENFEAT_pae_pgdir_above_4gb        4
+
+/* x86: Does this Xen host support the MMU_PT_UPDATE_PRESERVE_AD hypercall? */
+#define XENFEAT_mmu_pt_update_preserve_ad  5
+
+/* x86: Does this Xen host support the MMU_{CLEAR,COPY}_PAGE hypercall? */
+#define XENFEAT_highmem_assist             6
+
+/*
+ * If set, GNTTABOP_map_grant_ref honors flags to be placed into guest kernel
+ * available pte bits.
+ */
+#define XENFEAT_gnttab_map_avail_bits      7
+
+/* x86: Does this Xen host support the HVM callback vector type? */
+#define XENFEAT_hvm_callback_vector        8
+
+/* x86: pvclock algorithm is safe to use on HVM */
+#define XENFEAT_hvm_safe_pvclock           9
+
+/* x86: pirq can be used by HVM guests */
+#define XENFEAT_hvm_pirqs           10
+
+/* operation as Dom0 is supported */
+#define XENFEAT_dom0                      11
+
+/* Xen also maps grant references at pfn = mfn.
+ * This feature flag is deprecated and should not be used.
+#define XENFEAT_grant_map_identity        12
+ */
+
+/* Guest can use XENMEMF_vnode to specify virtual node for memory op. */
+#define XENFEAT_memory_op_vnode_supported 13
+
+/* arm: Hypervisor supports ARM SMC calling convention. */
+#define XENFEAT_ARM_SMCCC_supported       14
+
+/*
+ * x86/PVH: If set, ACPI RSDP can be placed at any address. Otherwise RSDP
+ * must be located in lower 1MB, as required by ACPI Specification for IA-PC
+ * systems.
+ * This feature flag is only consulted if XEN_ELFNOTE_GUEST_OS contains
+ * the "linux" string.
+ */
+#define XENFEAT_linux_rsdp_unrestricted   15
+
+/*
+ * A direct-mapped (or 1:1 mapped) domain is a domain for which its
+ * local pages have gfn == mfn. If a domain is direct-mapped,
+ * XENFEAT_direct_mapped is set; otherwise XENFEAT_not_direct_mapped
+ * is set.
+ *
+ * If neither flag is set (e.g. older Xen releases) the assumptions are:
+ * - not auto_translated domains (x86 only) are always direct-mapped
+ * - on x86, auto_translated domains are not direct-mapped
+ * - on ARM, Dom0 is direct-mapped, DomUs are not
+ */
+#define XENFEAT_not_direct_mapped         16
+#define XENFEAT_direct_mapped             17
+
+#define XENFEAT_NR_SUBMAPS 1
+
+#endif /* __XEN_PUBLIC_FEATURES_H__ */