tkernel: irqlatency: harden runtime and CPU hotplug state · Pull Request !988 · OpenCloudOS/OpenCloudOS-Kernel

## Summary

Harden the TKernel IRQ latency detector across enable/disable, CPU
hotplug, initialization failure, procfs reads, and module unload.

The change introduces one explicit runtime state machine, makes per-CPU
timer lifecycle operations idempotent, retains records for offline CPUs,
and extends the TKernel selftest with concurrent lifecycle coverage.

Related issue: #IKASKX

This is follow-up work to !974. It covers the complete detector lifecycle
and avoids wrapping cpuhp_remove_state() in an outer cpus_read_lock(),
because the hotplug removal helper already provides that serialization.

## Root cause

The detector publishes /proc/irq_latency before CPU hotplug registration
initializes the per-CPU timers. A concurrent enable can therefore start
uninitialized timer objects:

~~~
task 0                              task 1
------                              ------
create /proc/irq_latency
                                    write enable=1
                                    start per-CPU timers
register CPU hotplug callbacks
~~~

The timer callbacks also rearm unconditionally and have no explicit
per-CPU active state. Runtime mode changes, CPU online/offline, and module
exit are handled by separate paths, which makes duplicate starts and
shutdown ordering difficult to reason about.

In addition, trace_stack and trace_dist iterate a changing online CPU set.
Offline CPU history disappears from reports and cannot be cleared until
that CPU comes online again.

## Fix

- represent stopped, IRQ-running, softirq-running, and exiting modes with
  one state machine;
- serialize user-controlled transitions with control_lock and coordinate
  start/stop operations with CPU hotplug;
- add a per-CPU timers_active flag and stop both callbacks from rearming
  after disable, CPU offline, or module exit;
- initialize all possible CPUs before hotplug registration and publish
  procfs controls last;
- enter the exiting state before proc removal, remove the hotplug state,
  and only then free per-CPU storage;
- snapshot possible and online CPU masks under the CPU read lock, then
  release it before record copies and output formatting;
- retain and mark records from offline CPUs and clear all possible CPUs;
- add concurrent state-transition, CPU-hotplug, trace-read, and enabled
  module-unload selftests;
- document the lifecycle model and runtime impact.

The existing procfs ABI and enable values remain unchanged.

## Verification

Performed on WZU_Server using an x86_64 build:

- git diff --check: passed;
- strict checkpatch: 0 errors and 0 warnings;
- sh -n on the selftest: passed;
- full vmlinux build: passed;
- targeted W=1 irqlatency module build: passed;
- targeted sparse check: passed;
- bzImage and module build with lockdep and debug timers: passed;
- QEMU TKernel IRQ latency selftest: 15/15 passed;
- concurrent transitions, CPU offline/online, trace reads, and enabled
  module unload completed without lockdep, timer, or kernel warnings.

No kernel was installed, no boot configuration was changed, and no host
was rebooted.

类型	指派人员	状态
审查	zoucao Huang Cun Jianping Liu	已完成 (0/0人)
测试	Jianping Liu	已完成 (0/0人)

GVPOpenCloudOS / OpenCloudOS-Kernel

搜索帮助

GVP OpenCloudOS / OpenCloudOS-Kernel