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, 99 insertions, 0 deletions

diff --git a/Documentation/gpu/msm-preemption.rst b/Documentation/gpu/msm-preemption.rst
new file mode 100644
index 000000000..d768ca09f
--- /dev/null
+++ b/Documentation/gpu/msm-preemption.rst
@@ -0,0 +1,99 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+:orphan:
+
+==============
+MSM Preemption
+==============
+
+Preemption allows Adreno GPUs to switch to a higher priority ring when work is
+pushed to it, reducing latency for high priority submissions.
+
+When preemption is enabled 4 rings are initialized, corresponding to different
+priority levels. Having multiple rings is purely a software concept as the GPU
+only has registers to keep track of one graphics ring.
+The kernel is able to switch which ring is currently being processed by
+requesting preemption. When certain conditions are met, depending on the
+priority level, the GPU will save its current state in a series of buffers,
+then restores state from a similar set of buffers specified by the kernel. It
+then resumes execution and fires an IRQ to let the kernel know the context
+switch has completed.
+
+This mechanism can be used by the kernel to switch between rings. Whenever a
+submission occurs the kernel finds the highest priority ring which isn't empty
+and preempts to it if said ring is not the one being currently executed. This is
+also done whenever a submission completes to make sure execution resumes on a
+lower priority ring when a higher priority ring is done.
+
+Preemption levels
+-----------------
+
+Preemption can only occur at certain boundaries. The exact conditions can be
+configured by changing the preemption level, this allows to compromise between
+latency (ie. the time that passes between when the kernel requests preemption
+and when the SQE begins saving state) and overhead (the amount of state that
+needs to be saved).
+
+The GPU offers 3 levels:
+
+Level 0
+  Preemption only occurs at the submission level. This requires the least amount
+  of state to be saved as the execution of userspace submitted IBs is never
+  interrupted, however it offers very little benefit compared to not enabling
+  preemption of any kind.
+
+Level 1
+  Preemption occurs at either bin level, if using GMEM rendering, or draw level
+  in the sysmem rendering case.
+
+Level 2
+  Preemption occurs at draw level.
+
+Level 1 is the mode that is used by the msm driver.
+
+Additionally the GPU allows to specify a `skip_save_restore` option. This
+disables the saving and restoring of all registers except those relating to the
+operation of the SQE itself, reducing overhead. Saving and restoring is only
+skipped when using GMEM with Level 1 preemption. When enabling this userspace is
+expected to set the state that isn't preserved whenever preemption occurs which
+is done by specifying preamble and postambles. Those are IBs that are executed
+before and after preemption.
+
+Preemption buffers
+------------------
+
+A series of buffers are necessary to store the state of rings while they are not
+being executed. There are different kinds of preemption records and most of
+those require one buffer per ring. This is because preemption never occurs
+between submissions on the same ring, which always run in sequence when the ring
+is active. This means that only one context per ring is effectively active.
+
+SMMU_INFO
+  This buffer contains info about the current SMMU configuration such as the
+  ttbr0 register. The SQE firmware isn't actually able to save this record.
+  As a result SMMU info must be saved manually from the CP to a buffer and the
+  SMMU record updated with info from said buffer before triggering
+  preemption.
+
+NON_SECURE
+  This is the main preemption record where most state is saved. It is mostly
+  opaque to the kernel except for the first few words that must be initialized
+  by the kernel.
+
+SECURE
+  This saves state related to the GPU's secure mode.
+
+NON_PRIV
+  The intended purpose of this record is unknown. The SQE firmware actually
+  ignores it and therefore msm doesn't handle it.
+
+COUNTER
+  This record is used to save and restore performance counters.
+
+Handling the permissions of those buffers is critical for security. All but the
+NON_PRIV records need to be inaccessible from userspace, so they must be mapped
+in the kernel address space with the MSM_BO_MAP_PRIV flag.
+For example, making the NON_SECURE record accessible from userspace would allow
+any process to manipulate a saved ring's RPTR which can be used to skip the
+execution of some packets in a ring and execute user commands with higher
+privileges.