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, 82 insertions, 0 deletions

diff --git a/Documentation/virt/hyperv/clocks.rst b/Documentation/virt/hyperv/clocks.rst
new file mode 100644
index 000000000..176043265
--- /dev/null
+++ b/Documentation/virt/hyperv/clocks.rst
@@ -0,0 +1,82 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+Clocks and Timers
+=================
+
+arm64
+-----
+On arm64, Hyper-V virtualizes the ARMv8 architectural system counter
+and timer. Guest VMs use this virtualized hardware as the Linux
+clocksource and clockevents via the standard arm_arch_timer.c
+driver, just as they would on bare metal. Linux vDSO support for the
+architectural system counter is functional in guest VMs on Hyper-V.
+While Hyper-V also provides a synthetic system clock and four synthetic
+per-CPU timers as described in the TLFS, they are not used by the
+Linux kernel in a Hyper-V guest on arm64.  However, older versions
+of Hyper-V for arm64 only partially virtualize the ARMv8
+architectural timer, such that the timer does not generate
+interrupts in the VM. Because of this limitation, running current
+Linux kernel versions on these older Hyper-V versions requires an
+out-of-tree patch to use the Hyper-V synthetic clocks/timers instead.
+
+x86/x64
+-------
+On x86/x64, Hyper-V provides guest VMs with a synthetic system clock
+and four synthetic per-CPU timers as described in the TLFS. Hyper-V
+also provides access to the virtualized TSC via the RDTSC and
+related instructions. These TSC instructions do not trap to
+the hypervisor and so provide excellent performance in a VM.
+Hyper-V performs TSC calibration, and provides the TSC frequency
+to the guest VM via a synthetic MSR.  Hyper-V initialization code
+in Linux reads this MSR to get the frequency, so it skips TSC
+calibration and sets tsc_reliable. Hyper-V provides virtualized
+versions of the PIT (in Hyper-V  Generation 1 VMs only), local
+APIC timer, and RTC. Hyper-V does not provide a virtualized HPET in
+guest VMs.
+
+The Hyper-V synthetic system clock can be read via a synthetic MSR,
+but this access traps to the hypervisor. As a faster alternative,
+the guest can configure a memory page to be shared between the guest
+and the hypervisor.  Hyper-V populates this memory page with a
+64-bit scale value and offset value. To read the synthetic clock
+value, the guest reads the TSC and then applies the scale and offset
+as described in the Hyper-V TLFS. The resulting value advances
+at a constant 10 MHz frequency. In the case of a live migration
+to a host with a different TSC frequency, Hyper-V adjusts the
+scale and offset values in the shared page so that the 10 MHz
+frequency is maintained.
+
+Starting with Windows Server 2022 Hyper-V, Hyper-V uses hardware
+support for TSC frequency scaling to enable live migration of VMs
+across Hyper-V hosts where the TSC frequency may be different.
+When a Linux guest detects that this Hyper-V functionality is
+available, it prefers to use Linux's standard TSC-based clocksource.
+Otherwise, it uses the clocksource for the Hyper-V synthetic system
+clock implemented via the shared page (identified as
+"hyperv_clocksource_tsc_page").
+
+The Hyper-V synthetic system clock is available to user space via
+vDSO, and gettimeofday() and related system calls can execute
+entirely in user space.  The vDSO is implemented by mapping the
+shared page with scale and offset values into user space.  User
+space code performs the same algorithm of reading the TSC and
+applying the scale and offset to get the constant 10 MHz clock.
+
+Linux clockevents are based on Hyper-V synthetic timer 0 (stimer0).
+While Hyper-V offers 4 synthetic timers for each CPU, Linux only uses
+timer 0. In older versions of Hyper-V, an interrupt from stimer0
+results in a VMBus control message that is demultiplexed by
+vmbus_isr() as described in the Documentation/virt/hyperv/vmbus.rst
+documentation. In newer versions of Hyper-V, stimer0 interrupts can
+be mapped to an architectural interrupt, which is referred to as
+"Direct Mode". Linux prefers to use Direct Mode when available. Since
+x86/x64 doesn't support per-CPU interrupts, Direct Mode statically
+allocates an x86 interrupt vector (HYPERV_STIMER0_VECTOR) across all CPUs
+and explicitly codes it to call the stimer0 interrupt handler. Hence
+interrupts from stimer0 are recorded on the "HVS" line in /proc/interrupts
+rather than being associated with a Linux IRQ. Clockevents based on the
+virtualized PIT and local APIC timer also work, but Hyper-V stimer0
+is preferred.
+
+The driver for the Hyper-V synthetic system clock and timers is
+drivers/clocksource/hyperv_timer.c.