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, 224 insertions, 0 deletions

diff --git a/mm/fadvise.c b/mm/fadvise.c
new file mode 100644
index 000000000..b63fe2141
--- /dev/null
+++ b/mm/fadvise.c
@@ -0,0 +1,224 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * mm/fadvise.c
+ *
+ * Copyright (C) 2002, Linus Torvalds
+ *
+ * 11Jan2003	Andrew Morton
+ *		Initial version.
+ */
+
+#include <linux/kernel.h>
+#include <linux/file.h>
+#include <linux/fs.h>
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+#include <linux/backing-dev.h>
+#include <linux/fadvise.h>
+#include <linux/writeback.h>
+#include <linux/syscalls.h>
+#include <linux/swap.h>
+
+#include <asm/unistd.h>
+
+#include "internal.h"
+
+/*
+ * POSIX_FADV_WILLNEED could set PG_Referenced, and POSIX_FADV_NOREUSE could
+ * deactivate the pages and clear PG_Referenced.
+ */
+
+int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
+{
+	struct inode *inode;
+	struct address_space *mapping;
+	struct backing_dev_info *bdi;
+	loff_t endbyte;			/* inclusive */
+	pgoff_t start_index;
+	pgoff_t end_index;
+	unsigned long nrpages;
+
+	inode = file_inode(file);
+	if (S_ISFIFO(inode->i_mode))
+		return -ESPIPE;
+
+	mapping = file->f_mapping;
+	if (!mapping || len < 0 || offset < 0)
+		return -EINVAL;
+
+	bdi = inode_to_bdi(mapping->host);
+
+	if (IS_DAX(inode) || (bdi == &noop_backing_dev_info)) {
+		switch (advice) {
+		case POSIX_FADV_NORMAL:
+		case POSIX_FADV_RANDOM:
+		case POSIX_FADV_SEQUENTIAL:
+		case POSIX_FADV_WILLNEED:
+		case POSIX_FADV_NOREUSE:
+		case POSIX_FADV_DONTNEED:
+			/* no bad return value, but ignore advice */
+			break;
+		default:
+			return -EINVAL;
+		}
+		return 0;
+	}
+
+	/*
+	 * Careful about overflows. Len == 0 means "as much as possible".  Use
+	 * unsigned math because signed overflows are undefined and UBSan
+	 * complains.
+	 */
+	endbyte = (u64)offset + (u64)len;
+	if (!len || endbyte < len)
+		endbyte = LLONG_MAX;
+	else
+		endbyte--;		/* inclusive */
+
+	switch (advice) {
+	case POSIX_FADV_NORMAL:
+		file->f_ra.ra_pages = bdi->ra_pages;
+		spin_lock(&file->f_lock);
+		file->f_mode &= ~(FMODE_RANDOM | FMODE_NOREUSE);
+		spin_unlock(&file->f_lock);
+		break;
+	case POSIX_FADV_RANDOM:
+		spin_lock(&file->f_lock);
+		file->f_mode |= FMODE_RANDOM;
+		spin_unlock(&file->f_lock);
+		break;
+	case POSIX_FADV_SEQUENTIAL:
+		file->f_ra.ra_pages = bdi->ra_pages * 2;
+		spin_lock(&file->f_lock);
+		file->f_mode &= ~FMODE_RANDOM;
+		spin_unlock(&file->f_lock);
+		break;
+	case POSIX_FADV_WILLNEED:
+		/* First and last PARTIAL page! */
+		start_index = offset >> PAGE_SHIFT;
+		end_index = endbyte >> PAGE_SHIFT;
+
+		/* Careful about overflow on the "+1" */
+		nrpages = end_index - start_index + 1;
+		if (!nrpages)
+			nrpages = ~0UL;
+
+		force_page_cache_readahead(mapping, file, start_index, nrpages);
+		break;
+	case POSIX_FADV_NOREUSE:
+		spin_lock(&file->f_lock);
+		file->f_mode |= FMODE_NOREUSE;
+		spin_unlock(&file->f_lock);
+		break;
+	case POSIX_FADV_DONTNEED:
+		filemap_flush_range(mapping, offset, endbyte);
+
+		/*
+		 * First and last FULL page! Partial pages are deliberately
+		 * preserved on the expectation that it is better to preserve
+		 * needed memory than to discard unneeded memory.
+		 */
+		start_index = (offset+(PAGE_SIZE-1)) >> PAGE_SHIFT;
+		end_index = (endbyte >> PAGE_SHIFT);
+		/*
+		 * The page at end_index will be inclusively discarded according
+		 * by invalidate_mapping_pages(), so subtracting 1 from
+		 * end_index means we will skip the last page.  But if endbyte
+		 * is page aligned or is at the end of file, we should not skip
+		 * that page - discarding the last page is safe enough.
+		 */
+		if ((endbyte & ~PAGE_MASK) != ~PAGE_MASK &&
+				endbyte != inode->i_size - 1) {
+			/* First page is tricky as 0 - 1 = -1, but pgoff_t
+			 * is unsigned, so the end_index >= start_index
+			 * check below would be true and we'll discard the whole
+			 * file cache which is not what was asked.
+			 */
+			if (end_index == 0)
+				break;
+
+			end_index--;
+		}
+
+		if (end_index >= start_index) {
+			unsigned long nr_failed = 0;
+
+			/*
+			 * It's common to FADV_DONTNEED right after
+			 * the read or write that instantiates the
+			 * pages, in which case there will be some
+			 * sitting on the local LRU cache. Try to
+			 * avoid the expensive remote drain and the
+			 * second cache tree walk below by flushing
+			 * them out right away.
+			 */
+			lru_add_drain();
+
+			mapping_try_invalidate(mapping, start_index, end_index,
+					&nr_failed);
+
+			/*
+			 * The failures may be due to the folio being
+			 * in the LRU cache of a remote CPU. Drain all
+			 * caches and try again.
+			 */
+			if (nr_failed) {
+				lru_add_drain_all();
+				invalidate_mapping_pages(mapping, start_index,
+						end_index);
+			}
+		}
+		break;
+	default:
+		return -EINVAL;
+	}
+	return 0;
+}
+EXPORT_SYMBOL(generic_fadvise);
+
+int vfs_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
+{
+	if (file->f_op->fadvise)
+		return file->f_op->fadvise(file, offset, len, advice);
+
+	return generic_fadvise(file, offset, len, advice);
+}
+EXPORT_SYMBOL(vfs_fadvise);
+
+#ifdef CONFIG_ADVISE_SYSCALLS
+
+int ksys_fadvise64_64(int fd, loff_t offset, loff_t len, int advice)
+{
+	CLASS(fd, f)(fd);
+
+	if (fd_empty(f))
+		return -EBADF;
+
+	return vfs_fadvise(fd_file(f), offset, len, advice);
+}
+
+SYSCALL_DEFINE4(fadvise64_64, int, fd, loff_t, offset, loff_t, len, int, advice)
+{
+	return ksys_fadvise64_64(fd, offset, len, advice);
+}
+
+#ifdef __ARCH_WANT_SYS_FADVISE64
+
+SYSCALL_DEFINE4(fadvise64, int, fd, loff_t, offset, size_t, len, int, advice)
+{
+	return ksys_fadvise64_64(fd, offset, len, advice);
+}
+
+#endif
+
+#if defined(CONFIG_COMPAT) && defined(__ARCH_WANT_COMPAT_FADVISE64_64)
+
+COMPAT_SYSCALL_DEFINE6(fadvise64_64, int, fd, compat_arg_u64_dual(offset),
+		       compat_arg_u64_dual(len), int, advice)
+{
+	return ksys_fadvise64_64(fd, compat_arg_u64_glue(offset),
+				 compat_arg_u64_glue(len), advice);
+}
+
+#endif
+#endif