2023-06-20 14:38:42

by Waiman Long

Subject: [PATCH v2 5/5] x86/idle: Disable IBRS entering mwait idle and enable it on wakeup

When a user sets "intel_idle.max_cstate=0", it will disable
intel_idle and fall back to acpi_idle instead. The acpi_idle code
will then call mwait_idle_with_hints() to enter idle state. So when
X86_FEATURE_KERNEL_IBRS is enabled, it is necessary to disable IBRS
within mwait_idle_with_hints() when IRQ was disabled to avoid performance
degradation on silbing thread running user workload.

Signed-off-by: Waiman Long <[email protected]>
---
arch/x86/include/asm/mwait.h | 17 +++++++++++++++++
1 file changed, 17 insertions(+)

diff --git a/arch/x86/include/asm/mwait.h b/arch/x86/include/asm/mwait.h
index 778df05f8539..1e36cdc21661 100644
--- a/arch/x86/include/asm/mwait.h
+++ b/arch/x86/include/asm/mwait.h
@@ -108,15 +108,32 @@ static __always_inline void __sti_mwait(unsigned long eax, unsigned long ecx)
static __always_inline void mwait_idle_with_hints(unsigned long eax, unsigned long ecx)
{
if (static_cpu_has_bug(X86_BUG_MONITOR) || !current_set_polling_and_test()) {
+ bool ibrs_disabled = false;
+ u64 spec_ctrl;
+
if (static_cpu_has_bug(X86_BUG_CLFLUSH_MONITOR)) {
mb();
clflush((void *)&current_thread_info()->flags);
mb();
}

+ if (irqs_disabled() &&
+ cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS)) {
+ /* NMI always enable IBRS on exception entry */
+ ibrs_disabled = true;
+ spec_ctrl = spec_ctrl_current();
+ __this_cpu_write(x86_spec_ctrl_current, 0);
+ native_wrmsrl(MSR_IA32_SPEC_CTRL, 0);
+ }
+
__monitor((void *)&current_thread_info()->flags, 0, 0);
if (!need_resched())
__mwait(eax, ecx);
+
+ if (ibrs_disabled) {
+ native_wrmsrl(MSR_IA32_SPEC_CTRL, spec_ctrl);
+ __this_cpu_write(x86_spec_ctrl_current, spec_ctrl);
+ }
}
current_clr_polling();
}
--
2.31.1

2023-06-21 08:08:22

by Peter Zijlstra

Subject: Re: [PATCH v2 5/5] x86/idle: Disable IBRS entering mwait idle and enable it on wakeup

On Tue, Jun 20, 2023 at 10:06:25AM -0400, Waiman Long wrote:
> When a user sets "intel_idle.max_cstate=0", it will disable
> intel_idle and fall back to acpi_idle instead. The acpi_idle code
> will then call mwait_idle_with_hints() to enter idle state. So when
> X86_FEATURE_KERNEL_IBRS is enabled, it is necessary to disable IBRS
> within mwait_idle_with_hints() when IRQ was disabled to avoid performance
> degradation on silbing thread running user workload.

Urgh, no, just no. This is nasty.

2023-06-21 14:29:57

by Waiman Long

Subject: Re: [PATCH v2 5/5] x86/idle: Disable IBRS entering mwait idle and enable it on wakeup

On 6/21/23 03:32, Peter Zijlstra wrote:
> On Tue, Jun 20, 2023 at 10:06:25AM -0400, Waiman Long wrote:
>> When a user sets "intel_idle.max_cstate=0", it will disable
>> intel_idle and fall back to acpi_idle instead. The acpi_idle code
>> will then call mwait_idle_with_hints() to enter idle state. So when
>> X86_FEATURE_KERNEL_IBRS is enabled, it is necessary to disable IBRS
>> within mwait_idle_with_hints() when IRQ was disabled to avoid performance
>> degradation on silbing thread running user workload.
> Urgh, no, just no. This is nasty.

OK, will take this out in the next version.

Cheers,
Longman