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, 116 insertions, 0 deletions

diff --git a/Documentation/driver-api/driver-model/design-patterns.rst b/Documentation/driver-api/driver-model/design-patterns.rst
new file mode 100644
index 000000000..965b2b93b
--- /dev/null
+++ b/Documentation/driver-api/driver-model/design-patterns.rst
@@ -0,0 +1,116 @@
+=============================
+Device Driver Design Patterns
+=============================
+
+This document describes a few common design patterns found in device drivers.
+It is likely that subsystem maintainers will ask driver developers to
+conform to these design patterns.
+
+1. State Container
+2. container_of()
+
+
+1. State Container
+~~~~~~~~~~~~~~~~~~
+
+While the kernel contains a few device drivers that assume that they will
+only be probed() once on a certain system (singletons), it is custom to assume
+that the device the driver binds to will appear in several instances. This
+means that the probe() function and all callbacks need to be reentrant.
+
+The most common way to achieve this is to use the state container design
+pattern. It usually has this form::
+
+  struct foo {
+      spinlock_t lock; /* Example member */
+      (...)
+  };
+
+  static int foo_probe(...)
+  {
+      struct foo *foo;
+
+      foo = devm_kzalloc(dev, sizeof(*foo), GFP_KERNEL);
+      if (!foo)
+          return -ENOMEM;
+      spin_lock_init(&foo->lock);
+      (...)
+  }
+
+This will create an instance of struct foo in memory every time probe() is
+called. This is our state container for this instance of the device driver.
+Of course it is then necessary to always pass this instance of the
+state around to all functions that need access to the state and its members.
+
+For example, if the driver is registering an interrupt handler, you would
+pass around a pointer to struct foo like this::
+
+  static irqreturn_t foo_handler(int irq, void *arg)
+  {
+      struct foo *foo = arg;
+      (...)
+  }
+
+  static int foo_probe(...)
+  {
+      struct foo *foo;
+
+      (...)
+      ret = request_irq(irq, foo_handler, 0, "foo", foo);
+  }
+
+This way you always get a pointer back to the correct instance of foo in
+your interrupt handler.
+
+
+2. container_of()
+~~~~~~~~~~~~~~~~~
+
+Continuing on the above example we add an offloaded work::
+
+  struct foo {
+      spinlock_t lock;
+      struct workqueue_struct *wq;
+      struct work_struct offload;
+      (...)
+  };
+
+  static void foo_work(struct work_struct *work)
+  {
+      struct foo *foo = container_of(work, struct foo, offload);
+
+      (...)
+  }
+
+  static irqreturn_t foo_handler(int irq, void *arg)
+  {
+      struct foo *foo = arg;
+
+      queue_work(foo->wq, &foo->offload);
+      (...)
+  }
+
+  static int foo_probe(...)
+  {
+      struct foo *foo;
+
+      foo->wq = create_singlethread_workqueue("foo-wq");
+      INIT_WORK(&foo->offload, foo_work);
+      (...)
+  }
+
+The design pattern is the same for an hrtimer or something similar that will
+return a single argument which is a pointer to a struct member in the
+callback.
+
+container_of() is a macro defined in <linux/container_of.h>
+
+What container_of() does is to obtain a pointer to the containing struct from
+a pointer to a member by a simple subtraction using the offsetof() macro from
+standard C, which allows something similar to object oriented behaviours.
+Notice that the contained member must not be a pointer, but an actual member
+for this to work.
+
+We can see here that we avoid having global pointers to our struct foo *
+instance this way, while still keeping the number of parameters passed to the
+work function to a single pointer.