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, 173 insertions, 0 deletions

diff --git a/mm/vma_exec.c b/mm/vma_exec.c
new file mode 100644
index 000000000..586c52155
--- /dev/null
+++ b/mm/vma_exec.c
@@ -0,0 +1,173 @@
+// SPDX-License-Identifier: GPL-2.0-only
+
+/*
+ * Functions provided for exec functionality which however are
+ * specifically VMA-only logic.
+ */
+
+/*
+ * To allow for userland testing we place internal dependencies in
+ * vma_internal.h and external VMA API declarations in vma.h.
+ */
+#include "vma_internal.h"
+#include "vma.h"
+
+/*
+ * Relocate a VMA downwards by shift bytes. There cannot be any VMAs between
+ * this VMA and its relocated range, which will now reside at [vma->vm_start -
+ * shift, vma->vm_end - shift).
+ *
+ * This function is almost certainly NOT what you want for anything other than
+ * early executable temporary stack relocation.
+ */
+int relocate_vma_down(struct vm_area_struct *vma, unsigned long shift)
+{
+	/*
+	 * The process proceeds as follows:
+	 *
+	 * 1) Use shift to calculate the new vma endpoints.
+	 * 2) Extend vma to cover both the old and new ranges.  This ensures the
+	 *    arguments passed to subsequent functions are consistent.
+	 * 3) Move vma's page tables to the new range.
+	 * 4) Free up any cleared pgd range.
+	 * 5) Shrink the vma to cover only the new range.
+	 */
+
+	struct mm_struct *mm = vma->vm_mm;
+	unsigned long old_start = vma->vm_start;
+	unsigned long old_end = vma->vm_end;
+	unsigned long length = old_end - old_start;
+	unsigned long new_start = old_start - shift;
+	unsigned long new_end = old_end - shift;
+	VMA_ITERATOR(vmi, mm, new_start);
+	VMG_STATE(vmg, mm, &vmi, new_start, old_end, EMPTY_VMA_FLAGS,
+		  vma_start_pgoff(vma), vma_start_anon_pgoff(vma));
+	struct vm_area_struct *next;
+	struct mmu_gather tlb;
+	PAGETABLE_MOVE(pmc, vma, vma, old_start, new_start, length);
+
+	BUG_ON(new_start > new_end);
+
+	/*
+	 * ensure there are no vmas between where we want to go
+	 * and where we are
+	 */
+	if (vma != vma_next(&vmi))
+		return -EFAULT;
+
+	vma_iter_prev_range(&vmi);
+	/*
+	 * cover the whole range: [new_start, old_end)
+	 */
+	vmg.target = vma;
+	if (vma_expand(&vmg))
+		return -ENOMEM;
+
+	/*
+	 * move the page tables downwards, on failure we rely on
+	 * process cleanup to remove whatever mess we made.
+	 */
+	pmc.for_stack = true;
+	if (length != move_page_tables(&pmc))
+		return -ENOMEM;
+
+	tlb_gather_mmu(&tlb, mm);
+	next = vma_next(&vmi);
+	if (new_end > old_start) {
+		/*
+		 * when the old and new regions overlap clear from new_end.
+		 */
+		free_pgd_range(&tlb, new_end, old_end, new_end,
+			next ? next->vm_start : USER_PGTABLES_CEILING);
+	} else {
+		/*
+		 * otherwise, clean from old_start; this is done to not touch
+		 * the address space in [new_end, old_start) some architectures
+		 * have constraints on va-space that make this illegal (IA64) -
+		 * for the others its just a little faster.
+		 */
+		free_pgd_range(&tlb, old_start, old_end, new_end,
+			next ? next->vm_start : USER_PGTABLES_CEILING);
+	}
+	tlb_finish_mmu(&tlb);
+
+	vma_prev(&vmi);
+	/* Shrink the vma to just the new range */
+	return vma_shrink(&vmi, vma, new_end);
+}
+
+/*
+ * Establish the stack VMA in an execve'd process, located temporarily at the
+ * maximum stack address provided by the architecture.
+ *
+ * We later relocate this downwards in relocate_vma_down().
+ *
+ * This function is almost certainly NOT what you want for anything other than
+ * early executable initialisation.
+ *
+ * On success, returns 0 and sets *vmap to the stack VMA and *top_mem_p to the
+ * maximum addressable location in the stack (that is capable of storing a
+ * system word of data).
+ */
+int create_init_stack_vma(struct mm_struct *mm, struct vm_area_struct **vmap,
+			  unsigned long *top_mem_p)
+{
+	vma_flags_t flags = VMA_STACK_INCOMPLETE_SETUP;
+	struct vm_area_struct *vma;
+	int err;
+
+	/* VMA_STACK_FLAGS and VMA_STACK_INCOMPLETE_SETUP must not overlap. */
+	VM_WARN_ON_ONCE(vma_flags_test_any_mask(&flags, VMA_STACK_FLAGS));
+
+	vma = vm_area_alloc(mm);
+	if (!vma)
+		return -ENOMEM;
+
+	if (mmap_write_lock_killable(mm)) {
+		err = -EINTR;
+		goto err_free;
+	}
+
+	/*
+	 * Need to be called with mmap write lock
+	 * held, to avoid race with ksmd.
+	 */
+	err = ksm_execve(mm);
+	if (err)
+		goto err_ksm;
+
+	vma_flags_set_mask(&flags, VMA_STACK_FLAGS);
+	vma_set_anonymous(vma);
+
+	/*
+	 * Place the stack at the largest stack address the architecture
+	 * supports. Later, we'll move this to an appropriate place. We don't
+	 * use STACK_TOP because that can depend on attributes which aren't
+	 * configured yet.
+	 */
+	vma->vm_end = STACK_TOP_MAX;
+	vma->vm_start = vma->vm_end - PAGE_SIZE;
+	if (pgtable_supports_soft_dirty())
+		vma_flags_set(&flags, VMA_SOFTDIRTY_BIT);
+	vma->flags = flags;
+	vma->vm_page_prot = vma_get_page_prot(vma);
+
+	err = insert_vm_struct(mm, vma);
+	if (err)
+		goto err;
+
+	mm->stack_vm = mm->total_vm = 1;
+	mmap_write_unlock(mm);
+	*vmap = vma;
+	*top_mem_p = vma->vm_end - sizeof(void *);
+	return 0;
+
+err:
+	ksm_exit(mm);
+err_ksm:
+	mmap_write_unlock(mm);
+err_free:
+	*vmap = NULL;
+	vm_area_free(vma);
+	return err;
+}