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, 109 insertions, 0 deletions

diff --git a/Documentation/ABI/testing/debugfs-driver-qat b/Documentation/ABI/testing/debugfs-driver-qat
new file mode 100644
index 000000000..3f1efbbad
--- /dev/null
+++ b/Documentation/ABI/testing/debugfs-driver-qat
@@ -0,0 +1,109 @@
+What:		/sys/kernel/debug/qat_<device>_<BDF>/fw_counters
+Date:		November 2023
+KernelVersion:	6.6
+Contact:	qat-linux@intel.com
+Description:	(RO) Read returns the number of requests sent to the FW and the number of responses
+		received from the FW for each Acceleration Engine
+		Reported firmware counters::
+
+			<N>: Number of requests sent from Acceleration Engine N to FW and responses
+			     Acceleration Engine N received from FW
+
+What:		/sys/kernel/debug/qat_<device>_<BDF>/heartbeat/config
+Date:		November 2023
+KernelVersion:	6.6
+Contact:	qat-linux@intel.com
+Description:	(RW) Read returns value of the Heartbeat update period.
+		Write to the file changes this period value.
+
+		This period should reflect planned polling interval of device
+		health status. High frequency Heartbeat monitoring wastes CPU cycles
+		but minimizes the customer’s system downtime. Also, if there are
+		large service requests that take some time to complete, high frequency
+		Heartbeat monitoring could result in false reports of unresponsiveness
+		and in those cases, period needs to be increased.
+
+		This parameter is effective only for c3xxx, c62x, dh895xcc devices.
+		4xxx has this value internally fixed to 200ms.
+
+		Default value is set to 500. Minimal allowed value is 200.
+		All values are expressed in milliseconds.
+
+What:		/sys/kernel/debug/qat_<device>_<BDF>/heartbeat/queries_failed
+Date:		November 2023
+KernelVersion:	6.6
+Contact:	qat-linux@intel.com
+Description:	(RO) Read returns the number of times the device became unresponsive.
+
+		Attribute returns value of the counter which is incremented when
+		status query results negative.
+
+What:		/sys/kernel/debug/qat_<device>_<BDF>/heartbeat/queries_sent
+Date:		November 2023
+KernelVersion:	6.6
+Contact:	qat-linux@intel.com
+Description:	(RO) Read returns the number of times the control process checked
+		if the device is responsive.
+
+		Attribute returns value of the counter which is incremented on
+		every status query.
+
+What:		/sys/kernel/debug/qat_<device>_<BDF>/heartbeat/status
+Date:		November 2023
+KernelVersion:	6.6
+Contact:	qat-linux@intel.com
+Description:	(RO) Read returns the device health status.
+
+		Returns 0 when device is healthy or -1 when is unresponsive
+		or the query failed to send.
+
+		The driver does not monitor for Heartbeat. It is left for a user
+		to poll the status periodically.
+
+What:		/sys/kernel/debug/qat_<device>_<BDF>/pm_status
+Date:		January 2024
+KernelVersion:	6.7
+Contact:	qat-linux@intel.com
+Description:	(RO) Read returns power management information specific to the
+		QAT device.
+
+		This attribute is only available for qat_4xxx and qat_6xxx devices.
+
+What:		/sys/kernel/debug/qat_<device>_<BDF>/cnv_errors
+Date:		January 2024
+KernelVersion:	6.7
+Contact:	qat-linux@intel.com
+Description:	(RO) Read returns, for each Acceleration Engine (AE), the number
+		of errors and the type of the last error detected by the device
+		when performing verified compression.
+		Reported counters::
+
+			<N>: Number of Compress and Verify (CnV) errors and type
+			     of the last CnV error detected by Acceleration
+			     Engine N.
+
+What:		/sys/kernel/debug/qat_<device>_<BDF>/heartbeat/inject_error
+Date:		March 2024
+KernelVersion:	6.8
+Contact:	qat-linux@intel.com
+Description:	(WO) Write to inject an error that simulates an heartbeat
+		failure. This is to be used for testing purposes.
+
+		After writing this file, the driver stops arbitration on a
+		random engine and disables the fetching of heartbeat counters.
+		If a workload is running on the device, a job submitted to the
+		accelerator might not get a response and a read of the
+		`heartbeat/status` attribute might report -1, i.e. device
+		unresponsive.
+		The error is unrecoverable thus the device must be restarted to
+		restore its functionality.
+
+		This attribute is available only when the kernel is built with
+		CONFIG_CRYPTO_DEV_QAT_ERROR_INJECTION=y.
+
+		A write of 1 enables error injection.
+
+		The following example shows how to enable error injection::
+
+			# cd /sys/kernel/debug/qat_<device>_<BDF>
+			# echo 1 > heartbeat/inject_error