The Linux kernel provides mechanisms like 'isolcpus' and 'nohz_full' to
reduce interference for latency-sensitive workloads. However, these are
locked behind the "Reboot Wall" - they can only be configured via boot
parameters and require a system restart to change.
*** THIS IS AN RFC ***
This series is being submitted as an Request For Comments to discuss the
architectural changes required to support dynamic reconfiguration of
housekeeping boundaries. Key points for discussion:
- The use of a blocking notifier chain for cross-subsystem synchronization.
- Compatibility with existing boot-time isolcpus and nohz_full parameters.
- User-space sysfs interface design for granular isolation control.
While cgroup2 (cpuset) provides task-centric resource partitioning (including
unbound kthreads), it lacks the infrastructure to reconfigure core kernel
subsystems or manage per-CPU kernel activities. Specifically, cgroups cannot:
- Managed interrupt (IRQ) migration.
- RCU callback offloading and grace-period kthread placement.
- Per-CPU kernel threads (e.g., watchdog/n) and global subsystem logic (e.g., kcompactd).
- Subsystem-level masks for unbound workqueues.
- Dynamic toggling of full dynticks (NOHZ_FULL) mode.
This patch series introduces Dynamic Housekeeping & Enhanced Isolation
(DHEI). DHEI allows administrators to reconfigure the kernel's
housekeeping boundaries at runtime via a new sysfs interface at
/sys/kernel/housekeeping/.
Core Architecture:
1. Notifier-Driven Synchronization: A new blocking notifier chain
(HK_UPDATE_MASK) allows isolation.c to signal all participating
subsystems (IRQ, RCU, Sched, Watchdog, Workqueue, kcompactd, Tick)
whenever a housekeeping mask is modified.
2. Decoupled Memory Management: Replaced boot-time memory allocators
with runtime-safe variants, allowing masks to be allocated or resized
after the system is running.
3. Subsystem Handlers: Each critical subsystem implements a reconfiguration
handler to migrate pending work, re-affine kthreads, or re-route
interrupts dynamically.
Key Features:
- Fine-grained control: Separate sysfs nodes for timer, rcu, tick,
workqueue, kthread, etc.
- Dynamic NOHZ_FULL: Supports enabling/disabling full dynticks mode
on-the-fly by re-kicking CPUs to evaluate tick dependencies.
- SMT Awareness: An optional 'smt_aware_mode' ensures that all SMT
siblings of a physical core stay in the same isolation state.
- Safety Guard: Prevents the isolation of all CPUs, ensuring at least
one online CPU is always available for housekeeping tasks.
This series provides the necessary infrastructure for cloud-native
orchestrators and high-frequency trading platforms to dynamically
re-partition CPU resources without incurring the downtime of a reboot.
Signed-off-by: Qiliang Yuan <realwujing@gmail.com>
---
Qiliang Yuan (12):
sched/isolation: Remove __init restriction from housekeeping cores
sched/isolation: Introduce reconfiguration notifier chain
genirq: Implement dynamic migration for Managed IRQs
rcu: Sync RCU housekeeping mask on notification
sched/core: Dynamic update housekeeping_cpumask(HK_TYPE_DOMAIN)
watchdog: Allow runtime toggle of hardlockup detector on CPUs
workqueue: Dynamic housekeeping mask update support
kcompactd: Add housekeeping notifier for dynamic mask update
sched/isolation: Separate housekeeping types and add sysfs interface
tick/nohz: Implement dynamic nohz_full state update
sched/isolation: Implement SMT sibling auto-isolation and safety check
sched/isolation: Bridge isolcpus and support runtime tick offload init
include/linux/sched/isolation.h | 40 +++++--
include/linux/tick.h | 2 +-
kernel/irq/manage.c | 52 +++++++++
kernel/rcu/tree.c | 43 +++++++
kernel/sched/core.c | 5 +-
kernel/sched/isolation.c | 252 ++++++++++++++++++++++++++++++++++++++--
kernel/sched/sched.h | 2 +-
kernel/sched/topology.c | 26 +++++
kernel/time/tick-sched.c | 63 +++++++++-
kernel/watchdog.c | 24 ++++
kernel/workqueue.c | 39 +++++++
mm/compaction.c | 26 +++++
12 files changed, 547 insertions(+), 27 deletions(-)
---
base-commit: 1f97d9dcf53649c41c33227b345a36902cbb08ad
change-id: 20260206-feature-dynamic_isolcpus_dhei-ee46b6e3a477
Best regards,
--
Qiliang Yuan <realwujing@gmail.com>