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, 55 insertions, 0 deletions

diff --git a/Documentation/admin-guide/numastat.rst b/Documentation/admin-guide/numastat.rst
new file mode 100644
index 000000000..08ec2c2bd
--- /dev/null
+++ b/Documentation/admin-guide/numastat.rst
@@ -0,0 +1,55 @@
+===============================
+Numa policy hit/miss statistics
+===============================
+
+/sys/devices/system/node/node*/numastat
+
+All units are pages. Hugepages have separate counters.
+
+The numa_hit, numa_miss and numa_foreign counters reflect how well processes
+are able to allocate memory from nodes they prefer. If they succeed, numa_hit
+is incremented on the preferred node, otherwise numa_foreign is incremented on
+the preferred node and numa_miss on the node where allocation succeeded.
+
+Usually preferred node is the one local to the CPU where the process executes,
+but restrictions such as mempolicies can change that, so there are also two
+counters based on CPU local node. local_node is similar to numa_hit and is
+incremented on allocation from a node by CPU on the same node. other_node is
+similar to numa_miss and is incremented on the node where allocation succeeds
+from a CPU from a different node. Note there is no counter analogical to
+numa_foreign.
+
+In more detail:
+
+=============== ============================================================
+numa_hit	A process wanted to allocate memory from this node,
+		and succeeded.
+
+numa_miss	A process wanted to allocate memory from another node,
+		but ended up with memory from this node.
+
+numa_foreign	A process wanted to allocate on this node,
+		but ended up with memory from another node.
+
+local_node	A process ran on this node's CPU,
+		and got memory from this node.
+
+other_node	A process ran on a different node's CPU
+		and got memory from this node.
+
+interleave_hit 	Interleaving wanted to allocate from this node
+		and succeeded.
+=============== ============================================================
+
+For easier reading you can use the numastat utility from the numactl package
+(http://oss.sgi.com/projects/libnuma/). Note that it only works
+well right now on machines with a small number of CPUs.
+
+Note that on systems with memoryless nodes (where a node has CPUs but no
+memory) the numa_hit, numa_miss and numa_foreign statistics can be skewed
+heavily. In the current kernel implementation, if a process prefers a
+memoryless node (i.e.  because it is running on one of its local CPU), the
+implementation actually treats one of the nearest nodes with memory as the
+preferred node. As a result, such allocation will not increase the numa_foreign
+counter on the memoryless node, and will skew the numa_hit, numa_miss and
+numa_foreign statistics of the nearest node.