On 2025/11/19 上午10:58, Huacai Chen wrote:
On Wed, Nov 19, 2025 at 10:01 AM Bibo Mao <[email protected]> wrote:



On 2025/11/18 下午8:48, Huacai Chen wrote:
Hi, Bibo,

On Tue, Nov 18, 2025 at 4:07 PM Bibo Mao <[email protected]> wrote:

Function vcpu_is_preempted() is used to check whether vCPU is preempted
or not. Here add implementation with vcpu_is_preempted() when option
CONFIG_PARAVIRT is enabled.

Signed-off-by: Bibo Mao <[email protected]>
---
   arch/loongarch/include/asm/smp.h      |  1 +
   arch/loongarch/include/asm/spinlock.h |  5 +++++
   arch/loongarch/kernel/paravirt.c      | 16 ++++++++++++++++
   arch/loongarch/kernel/smp.c           |  6 ++++++
   4 files changed, 28 insertions(+)

diff --git a/arch/loongarch/include/asm/smp.h b/arch/loongarch/include/asm/smp.h
index 3a47f52959a8..5b37f7bf2060 100644
--- a/arch/loongarch/include/asm/smp.h
+++ b/arch/loongarch/include/asm/smp.h
@@ -18,6 +18,7 @@ struct smp_ops {
          void (*init_ipi)(void);
          void (*send_ipi_single)(int cpu, unsigned int action);
          void (*send_ipi_mask)(const struct cpumask *mask, unsigned int 
action);
+       bool (*vcpu_is_preempted)(int cpu);
   };
   extern struct smp_ops mp_ops;

diff --git a/arch/loongarch/include/asm/spinlock.h 
b/arch/loongarch/include/asm/spinlock.h
index 7cb3476999be..c001cef893aa 100644
--- a/arch/loongarch/include/asm/spinlock.h
+++ b/arch/loongarch/include/asm/spinlock.h
@@ -5,6 +5,11 @@
   #ifndef _ASM_SPINLOCK_H
   #define _ASM_SPINLOCK_H

+#ifdef CONFIG_PARAVIRT
+#define vcpu_is_preempted      vcpu_is_preempted
+bool vcpu_is_preempted(int cpu);
+#endif
Maybe paravirt.h is a better place?

It is actually a little strange to add macro CONFIG_PARAVIRT in file
asm/spinlock.h

vcpu_is_preempted is originally defined in header file
include/linux/sched.h like this
#ifndef vcpu_is_preempted
static inline bool vcpu_is_preempted(int cpu)
{
          return false;
}
#endif

that requires that header file is included before sched.h, file
asm/spinlock.h can meet this requirement, however header file paravirt.h
maybe it is not included before sched.h in generic.

Here vcpu_is_preempted definition is added before the following including.
     #include <asm/processor.h>
     #include <asm/qspinlock.h>
     #include <asm/qrwlock.h>
Maybe it is better to be added after the above header files including
sentences, but need further investigation.
powerpc put it in paravirt.h, so I think it is possible.
paravirt.h is included by header file asm/qspinlock.h on powerpc, however it is not so on loongarch :)

# grep paravirt.h arch/powerpc/* -r
arch/powerpc/include/asm/paravirt_api_clock.h:#include <asm/paravirt.h>
arch/powerpc/include/asm/qspinlock.h:#include <asm/paravirt.h>
arch/powerpc/include/asm/simple_spinlock.h:#include <asm/paravirt.h>

$ grep paravirt.h arch/loongarch/* -r
arch/loongarch/include/asm/paravirt_api_clock.h:#include <asm/paravirt.h>


+
   #include <asm/processor.h>
   #include <asm/qspinlock.h>
   #include <asm/qrwlock.h>
diff --git a/arch/loongarch/kernel/paravirt.c b/arch/loongarch/kernel/paravirt.c
index b1b51f920b23..b99404b6b13f 100644
--- a/arch/loongarch/kernel/paravirt.c
+++ b/arch/loongarch/kernel/paravirt.c
@@ -52,6 +52,13 @@ static u64 paravt_steal_clock(int cpu)
   #ifdef CONFIG_SMP
   static struct smp_ops native_ops;

+static bool pv_vcpu_is_preempted(int cpu)
+{
+       struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
+
+       return !!(src->preempted & KVM_VCPU_PREEMPTED);
+}
+
   static void pv_send_ipi_single(int cpu, unsigned int action)
   {
          int min, old;
@@ -308,6 +315,9 @@ int __init pv_time_init(void)
                  pr_err("Failed to install cpu hotplug callbacks\n");
                  return r;
          }
+
+       if (kvm_para_has_feature(KVM_FEATURE_PREEMPT_HINT))
+               mp_ops.vcpu_is_preempted = pv_vcpu_is_preempted;
   #endif

          static_call_update(pv_steal_clock, paravt_steal_clock);
@@ -332,3 +342,9 @@ int __init pv_spinlock_init(void)

          return 0;
   }
+
+bool notrace vcpu_is_preempted(int cpu)
+{
+       return mp_ops.vcpu_is_preempted(cpu);
+}

We can simplify the whole patch like this, then we don't need to touch
smp.c, and we can merge Patch-2/3.

+bool notrace vcpu_is_preempted(int cpu)
+{
+  if (!kvm_para_has_feature(KVM_FEATURE_PREEMPT_HINT))
+     return false;
+ else {
+     struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
+     return !!(src->preempted & KVM_VCPU_PREEMPTED);
+ }
+}
1. there is assembly output about relative vcpu_is_preempted
   <loongson_vcpu_is_preempted>:
                 move    $r4,$r0
                 jirl    $r0,$r1,0

   <pv_vcpu_is_preempted>:
          pcalau12i       $r13,8759(0x2237)
          slli.d  $r4,$r4,0x3
          addi.d  $r13,$r13,-1000(0xc18)
          ldx.d   $r13,$r13,$r4
          pcalau12i       $r12,5462(0x1556)
          addi.d  $r12,$r12,384(0x180)
          add.d   $r12,$r13,$r12
          ld.bu   $r4,$r12,16(0x10)
          andi    $r4,$r4,0x1
          jirl    $r0,$r1,0

   <vcpu_is_preempted>:
          pcalau12i       $r12,8775(0x2247)
          ld.d    $r12,$r12,-472(0xe28)
          jirl    $r0,$r12,0
          andi    $r0,$r0,0x0

   <vcpu_is_preempted_new>:
          pcalau12i       $r12,8151(0x1fd7)
          ld.d    $r12,$r12,-1008(0xc10)
          bstrpick.d      $r12,$r12,0x1a,0x1a
          beqz    $r12,188(0xbc) # 900000000024ec60
          pcalau12i       $r12,11802(0x2e1a)
          addi.d  $r12,$r12,-1400(0xa88)
          ldptr.w $r14,$r12,36(0x24)
          beqz    $r14,108(0x6c) # 900000000024ec20
          addi.w  $r13,$r0,1(0x1)
          bne     $r14,$r13,164(0xa4) # 900000000024ec60
          ldptr.w $r13,$r12,40(0x28)
          bnez    $r13,24(0x18) # 900000000024ebdc
          lu12i.w $r14,262144(0x40000)
          ori     $r14,$r14,0x4
          cpucfg  $r14,$r14
          slli.w  $r13,$r14,0x0
          st.w    $r14,$r12,40(0x28)
          bstrpick.d      $r13,$r13,0x3,0x3
          beqz    $r13,128(0x80) # 900000000024ec60
          pcalau12i       $r13,8759(0x2237)
          slli.d  $r4,$r4,0x3
          addi.d  $r13,$r13,-1000(0xc18)
          ldx.d   $r13,$r13,$r4
          pcalau12i       $r12,5462(0x1556)
          addi.d  $r12,$r12,384(0x180)
          add.d   $r12,$r13,$r12
          ld.bu   $r4,$r12,16(0x10)
          andi    $r4,$r4,0x1
          jirl    $r0,$r1,0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          lu12i.w $r13,262144(0x40000)
          cpucfg  $r13,$r13
          lu12i.w $r15,1237(0x4d5)
          ori     $r15,$r15,0x64b
          slli.w  $r13,$r13,0x0
          bne     $r13,$r15,-124(0x3ff84) # 900000000024ebb8
          addi.w  $r13,$r0,1(0x1)
          st.w    $r13,$r12,36(0x24)
          b       -128(0xfffff80) # 900000000024ebc0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          andi    $r0,$r0,0x0
          move    $r4,$r0
          jirl    $r0,$r1,0

With vcpu_is_preempted(), there is one memory load and one jirl jump,
with vcpu_is_preempted_new(), there is two memory load and two beq
compare instructions.
Is vcpu_is_preempted() performance critical (we need performance data
here)? It seems the powerpc version is also complex.


2. In some scenery such nr_cpus == 1, loongson_vcpu_is_preempted() is
better than pv_vcpu_is_preempted() even if the preempt feature is enabled.
In your original patch, "mp_ops.vcpu_is_preempted =
pv_vcpu_is_preempted" if the preempt feature is enabled. Why is
loongson_vcpu_is_preempted() called when nr_cpus=1?

Huacai


Regards
Bibo Mao
Huacai

+EXPORT_SYMBOL(vcpu_is_preempted);
diff --git a/arch/loongarch/kernel/smp.c b/arch/loongarch/kernel/smp.c
index 46036d98da75..f04192fedf8d 100644
--- a/arch/loongarch/kernel/smp.c
+++ b/arch/loongarch/kernel/smp.c
@@ -307,10 +307,16 @@ static void loongson_init_ipi(void)
                  panic("IPI IRQ request failed\n");
   }

+static bool loongson_vcpu_is_preempted(int cpu)
+{
+       return false;
+}
+
   struct smp_ops mp_ops = {
          .init_ipi               = loongson_init_ipi,
          .send_ipi_single        = loongson_send_ipi_single,
          .send_ipi_mask          = loongson_send_ipi_mask,
+       .vcpu_is_preempted      = loongson_vcpu_is_preempted,
   };

   static void __init fdt_smp_setup(void)
--
2.39.3






Reply via email to