Re: [PATCH] LoongArch: time: Fix the issue of high cpu usage of vcpu threads in virtual machines

From: lixianglai
Date: Tue Aug 12 2025 - 08:50:58 EST




Hi Huacai Chen:
Hi, Xianglai,

There is something that can be improved.

On Tue, Aug 12, 2025 at 5:24 PM Xianglai Li <lixianglai@xxxxxxxxxxx> wrote:
When the cpu is offline, the timer under loongarch is not correctly closed,
resulting in an excessively high cpu usage rate of the offline vcpu thread
in the virtual machine.

To correctly close the timer, we have made the following modifications:

Register the cpu hotplug timer start event for loongarch.This event will
be called to close the timer when the cpu is offline.

Clear the timer interrupt when the timer is turned off

Signed-off-by: Xianglai Li <lixianglai@xxxxxxxxxxx>
---
arch/loongarch/kernel/time.c | 20 ++++++++++++++++++++
include/linux/cpuhotplug.h | 1 +
2 files changed, 21 insertions(+)

diff --git a/arch/loongarch/kernel/time.c b/arch/loongarch/kernel/time.c
index 367906b10f81..4daa11512eba 100644
--- a/arch/loongarch/kernel/time.c
+++ b/arch/loongarch/kernel/time.c
@@ -12,6 +12,7 @@
#include <linux/kernel.h>
#include <linux/sched_clock.h>
#include <linux/spinlock.h>
+#include <linux/cpu.h>

#include <asm/cpu-features.h>
#include <asm/loongarch.h>
@@ -86,6 +87,9 @@ static int constant_set_state_shutdown(struct clock_event_device *evt)
timer_config &= ~CSR_TCFG_EN;
csr_write64(timer_config, LOONGARCH_CSR_TCFG);

+ /* Clear Timer Interrupt */
+ write_csr_tintclear(CSR_TINTCLR_TI);
+
raw_spin_unlock(&state_lock);

return 0;
@@ -208,8 +212,17 @@ int __init constant_clocksource_init(void)
return res;
}

+static int arch_timer_dying_cpu(unsigned int cpu)
We can use arch_timer_dying() for short. And then add an
arch_timer_starting() like this:

static int arch_timer_starting(unsigned int cpu)
{
set_csr_ecfg(ECFGF_TIMER);

return 0;
}

Though ECFGF_TIMER may be enabled in other places, for syntax we need it here.

+{
+ constant_set_state_shutdown(NULL);
+
+ return 0;
+}
+
void __init time_init(void)
{
+ int err;
+
if (!cpu_has_cpucfg)
const_clock_freq = cpu_clock_freq;
else
@@ -220,4 +233,11 @@ void __init time_init(void)
constant_clockevent_init();
constant_clocksource_init();
pv_time_init();
+
+ err = cpuhp_setup_state_nocalls(CPUHP_AP_LOONGARCH_ARCH_TIMER_STARTING,
+ "loongarch/timer:starting",
+ NULL, arch_timer_dying_cpu);
Then we need use cpuhp_setup_state() here, because we have a startup
function now.

And "loongarch/timer:starting" should be
"clockevents/loongarch/timer:starting" like others.

And the whole should be moved to the last of
constant_clockevent_init() because it is clockevent specific.

like this?

@@ -164,6 +182,10 @@ int constant_clockevent_init(void)

        timer_irq_installed = 1;

+       cpuhp_setup_state(CPUHP_AP_LOONGARCH_ARCH_TIMER_STARTING,
+                         "clockevents/loongarch/timer:starting",
+                         arch_timer_starting, arch_timer_dying_cpu);
+
        sync_counter();


I was wondering whether it should be placed before or after the "timer_irq_installed" judgment


+ if (err)
+ pr_info("cpu hotplug event register failed");
This is not so useful, because the error isn't fatal.


Huacai

+
}
diff --git a/include/linux/cpuhotplug.h b/include/linux/cpuhotplug.h
index edfa61d80702..6606c1546afc 100644
--- a/include/linux/cpuhotplug.h
+++ b/include/linux/cpuhotplug.h
@@ -159,6 +159,7 @@ enum cpuhp_state {
CPUHP_AP_PERF_ARM_STARTING,
CPUHP_AP_PERF_RISCV_STARTING,
CPUHP_AP_ARM_L2X0_STARTING,
+ CPUHP_AP_LOONGARCH_ARCH_TIMER_STARTING,
CPUHP_AP_EXYNOS4_MCT_TIMER_STARTING,
CPUHP_AP_ARM_ARCH_TIMER_STARTING,
CPUHP_AP_ARM_ARCH_TIMER_EVTSTRM_STARTING,

base-commit: 53e760d8949895390e256e723e7ee46618310361
--
2.39.1