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

2 files changed, 92 insertions, 0 deletions

diff --git a/Documentation/gpu/amdgpu/gc/index.rst b/Documentation/gpu/amdgpu/gc/index.rst
new file mode 100644
index 000000000..b6b583c9d
--- /dev/null
+++ b/Documentation/gpu/amdgpu/gc/index.rst
@@ -0,0 +1,54 @@
+.. _amdgpu-gc:
+
+========================================
+ drm/amdgpu - Graphics and Compute (GC)
+========================================
+
+The relationship between the CPU and GPU can be described as the
+producer-consumer problem, where the CPU fills out a buffer with operations
+(producer) to be executed by the GPU (consumer). The requested operations in
+the buffer are called **Command Packets**, which can be summarized as a
+compressed way of transmitting command information to the graphics controller.
+
+The component that acts as the front end between the CPU and the GPU is called
+**Command Processor (CP)**. This component is responsible for providing greater
+flexibility to the **Graphics and Compute (GC)** since CP makes it possible to
+program various aspects of the GPU pipeline. CP also coordinates the
+communication between the CPU and GPU via a mechanism named **Ring Buffers**,
+where the CPU appends information to the buffer while the GPU removes
+operations. CP is also responsible for handling **Indirect Buffers (IB)**.
+
+For reference, internally the CP consists of several sub-blocks (CPC - CP
+compute, CPG - CP graphics, and CPF - CP fetcher). Some of these acronyms
+appear in register names, but this is more of an implementation detail and not
+something that directly impacts driver programming or debugging.
+
+Graphics (GFX) and Compute Microcontrollers
+-------------------------------------------
+
+GC is a large block, and as a result, it has multiple firmware associated with
+it. Some of them are:
+
+CP (Command Processor)
+    The name for the hardware block that encompasses the front end of the
+    GFX/Compute pipeline. Consists mainly of a bunch of microcontrollers
+    (PFP, ME, CE, MEC). The firmware that runs on these microcontrollers
+    provides the driver interface to interact with the GFX/Compute engine.
+
+    MEC (MicroEngine Compute)
+        This is the microcontroller that controls the compute queues on the
+        GFX/compute engine.
+
+    MES (MicroEngine Scheduler)
+        This is the engine for managing queues. For more details check
+        :ref:`MicroEngine Scheduler (MES) <amdgpu-mes>`.
+
+RLC (RunList Controller)
+    This is another microcontroller in the GFX/Compute engine. It handles
+    power management related functionality within the GFX/Compute engine.
+    The name is a vestige of old hardware where it was originally added
+    and doesn't really have much relation to what the engine does now.
+
+.. toctree::
+
+   mes.rst
diff --git a/Documentation/gpu/amdgpu/gc/mes.rst b/Documentation/gpu/amdgpu/gc/mes.rst
new file mode 100644
index 000000000..b99eb211b
--- /dev/null
+++ b/Documentation/gpu/amdgpu/gc/mes.rst
@@ -0,0 +1,38 @@
+.. _amdgpu-mes:
+
+=============================
+ MicroEngine Scheduler (MES)
+=============================
+
+.. note::
+   Queue and ring buffer are used as a synonymous.
+
+.. note::
+   This section assumes that you are familiar with the concept of Pipes, Queues, and GC.
+   If not, check :ref:`GFX, Compute, and SDMA Overall Behavior<pipes-and-queues-description>`
+   and :ref:`drm/amdgpu - Graphics and Compute (GC) <amdgpu-gc>`.
+
+Every GFX has a pipe component with one or more hardware queues. Pipes can
+switch between queues depending on certain conditions, and one of the
+components that can request a queue switch to a pipe is the MicroEngine
+Scheduler (MES). Whenever the driver is initialized, it creates one MQD per
+hardware queue, and then the MQDs are handed to the MES firmware for mapping
+to:
+
+1. Kernel Queues (legacy): This queue is statically mapped to HQDs and never
+   preempted. Even though this is a legacy feature, it is the current default, and
+   most existing hardware supports it. When an application submits work to the
+   kernel driver, it submits all of the application command buffers to the kernel
+   queues. The CS IOCTL takes the command buffer from the applications and
+   schedules them on the kernel queue.
+
+2. User Queues: These queues are dynamically mapped to the HQDs. Regarding the
+   utilization of User Queues, the userspace application will create its user
+   queues and submit work directly to its user queues with no need to IOCTL for
+   each submission and no need to share a single kernel queue.
+
+In terms of User Queues, MES can dynamically map them to the HQD. If there are
+more MQDs than HQDs, the MES firmware will preempt other user queues to make
+sure each queues get a time slice; in other words, MES is a microcontroller
+that handles the mapping and unmapping of MQDs into HQDs, as well as the
+priorities and oversubscription of MQDs.