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, 157 insertions, 0 deletions

diff --git a/Documentation/ABI/testing/debugfs-cxl b/Documentation/ABI/testing/debugfs-cxl
new file mode 100644
index 000000000..2989d4da9
--- /dev/null
+++ b/Documentation/ABI/testing/debugfs-cxl
@@ -0,0 +1,157 @@
+What:		/sys/kernel/debug/cxl/memX/inject_poison
+Date:		April, 2023
+KernelVersion:	v6.4
+Contact:	linux-cxl@vger.kernel.org
+Description:
+		(WO) When a Device Physical Address (DPA) is written to this
+		attribute, the memdev driver sends an inject poison command to
+		the device for the specified address. The DPA must be 64-byte
+		aligned and the length of the injected poison is 64-bytes. If
+		successful, the device returns poison when the address is
+		accessed through the CXL.mem bus. Injecting poison adds the
+		address to the device's Poison List and the error source is set
+		to Injected. In addition, the device adds a poison creation
+		event to its internal Informational Event log, updates the
+		Event Status register, and if configured, interrupts the host.
+		It is not an error to inject poison into an address that
+		already has poison present and no error is returned. If the
+		device returns 'Inject Poison Limit Reached' an -EBUSY error
+		is returned to the user. The inject_poison attribute is only
+		visible for devices supporting the capability.
+
+		TEST-ONLY INTERFACE: This interface is intended for testing
+		and validation purposes only. It is not a data repair mechanism
+		and should never be used on production systems or live data.
+
+		DATA LOSS RISK: For CXL persistent memory (PMEM) devices,
+		poison injection can result in permanent data loss. Injected
+		poison may render data permanently inaccessible even after
+		clearing, as the clear operation writes zeros and does not
+		recover original data.
+
+		SYSTEM STABILITY RISK: For volatile memory, poison injection
+		can cause kernel crashes, system instability, or unpredictable
+		behavior if the poisoned addresses are accessed by running code
+		or critical kernel structures.
+
+What:		/sys/kernel/debug/cxl/memX/clear_poison
+Date:		April, 2023
+KernelVersion:	v6.4
+Contact:	linux-cxl@vger.kernel.org
+Description:
+		(WO) When a Device Physical Address (DPA) is written to this
+		attribute, the memdev driver sends a clear poison command to
+		the device for the specified address. Clearing poison removes
+		the address from the device's Poison List and writes 0 (zero)
+		for 64 bytes starting at address. It is not an error to clear
+		poison from an address that does not have poison set. If the
+		device cannot clear poison from the address, -ENXIO is returned.
+		The clear_poison attribute is only visible for devices
+		supporting the capability.
+
+		TEST-ONLY INTERFACE: This interface is intended for testing
+		and validation purposes only. It is not a data repair mechanism
+		and should never be used on production systems or live data.
+
+		CLEAR IS NOT DATA RECOVERY: This operation writes zeros to the
+		specified address range and removes the address from the poison
+		list. It does NOT recover or restore original data that may have
+		been present before poison injection. Any original data at the
+		cleared address is permanently lost and replaced with zeros.
+
+		CLEAR IS NOT A REPAIR MECHANISM: This interface is for testing
+		purposes only and should not be used as a data repair tool.
+		Clearing poison is fundamentally different from data recovery
+		or error correction.
+
+What:		/sys/kernel/debug/cxl/regionX/inject_poison
+Date:		August, 2025
+Contact:	linux-cxl@vger.kernel.org
+Description:
+		(WO) When a Host Physical Address (HPA) is written to this
+		attribute, the region driver translates it to a Device
+		Physical Address (DPA) and identifies the corresponding
+		memdev. It then sends an inject poison command to that memdev
+		at the translated DPA. Refer to the memdev ABI entry at:
+		/sys/kernel/debug/cxl/memX/inject_poison for the detailed
+		behavior. This attribute is only visible if all memdevs
+		participating in the region support both inject and clear
+		poison commands.
+
+		TEST-ONLY INTERFACE: This interface is intended for testing
+		and validation purposes only. It is not a data repair mechanism
+		and should never be used on production systems or live data.
+
+		DATA LOSS RISK: For CXL persistent memory (PMEM) devices,
+		poison injection can result in permanent data loss. Injected
+		poison may render data permanently inaccessible even after
+		clearing, as the clear operation writes zeros and does not
+		recover original data.
+
+		SYSTEM STABILITY RISK: For volatile memory, poison injection
+		can cause kernel crashes, system instability, or unpredictable
+		behavior if the poisoned addresses are accessed by running code
+		or critical kernel structures.
+
+What:		/sys/kernel/debug/cxl/regionX/clear_poison
+Date:		August, 2025
+Contact:	linux-cxl@vger.kernel.org
+Description:
+		(WO) When a Host Physical Address (HPA) is written to this
+		attribute, the region driver translates it to a Device
+		Physical Address (DPA) and identifies the corresponding
+		memdev. It then sends a clear poison command to that memdev
+		at the translated DPA. Refer to the memdev ABI entry at:
+		/sys/kernel/debug/cxl/memX/clear_poison for the detailed
+		behavior. This attribute is only visible if all memdevs
+		participating in the region support both inject and clear
+		poison commands.
+
+		TEST-ONLY INTERFACE: This interface is intended for testing
+		and validation purposes only. It is not a data repair mechanism
+		and should never be used on production systems or live data.
+
+		CLEAR IS NOT DATA RECOVERY: This operation writes zeros to the
+		specified address range and removes the address from the poison
+		list. It does NOT recover or restore original data that may have
+		been present before poison injection. Any original data at the
+		cleared address is permanently lost and replaced with zeros.
+
+		CLEAR IS NOT A REPAIR MECHANISM: This interface is for testing
+		purposes only and should not be used as a data repair tool.
+		Clearing poison is fundamentally different from data recovery
+		or error correction.
+
+What:		/sys/kernel/debug/cxl/einj_types
+Date:		January, 2024
+KernelVersion:	v6.9
+Contact:	linux-cxl@vger.kernel.org
+Description:
+		(RO) Prints the CXL protocol error types made available by
+		the platform in the format:
+
+			0x<error number> <error type>
+
+		The possible error types are (as of ACPI v6.5):
+
+			0x1000	CXL.cache Protocol Correctable
+			0x2000	CXL.cache Protocol Uncorrectable non-fatal
+			0x4000	CXL.cache Protocol Uncorrectable fatal
+			0x8000	CXL.mem Protocol Correctable
+			0x10000	CXL.mem Protocol Uncorrectable non-fatal
+			0x20000	CXL.mem Protocol Uncorrectable fatal
+
+		The <error number> can be written to einj_inject to inject
+		<error type> into a chosen dport.
+
+What:		/sys/kernel/debug/cxl/$dport_dev/einj_inject
+Date:		January, 2024
+KernelVersion:	v6.9
+Contact:	linux-cxl@vger.kernel.org
+Description:
+		(WO) Writing an integer to this file injects the corresponding
+		CXL protocol error into $dport_dev ($dport_dev will be a device
+		name from /sys/bus/pci/devices). The integer to type mapping for
+		injection can be found by reading from einj_types. If the dport
+		was enumerated in RCH mode, a CXL 1.1 error is injected, otherwise
+		a CXL 2.0 error is injected.