Re: [PATCH] smp/hotplug, x86/vmware: Put offline vCPUs in halt instead of mwait

[Juergen Gross <jgross@suse.com>]

[-- Attachment #1.1.1: Type: text/plain, Size: 3200 bytes --]

On 23.09.22 09:05, Peter Zijlstra wrote:
> On Thu, Jul 21, 2022 at 01:44:33PM -0700, Srivatsa S. Bhat wrote:
>> From: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
>>
>> VMware ESXi allows enabling a passthru mwait CPU-idle state in the
>> guest using the following VMX option:
>>
>> monitor_control.mwait_in_guest = "TRUE"
>>
>> This lets a vCPU in mwait to remain in guest context (instead of
>> yielding to the hypervisor via a VMEXIT), which helps speed up
>> wakeups from idle.
>>
>> However, this runs into problems with CPU hotplug, because the Linux
>> CPU offline path prefers to put the vCPU-to-be-offlined in mwait
>> state, whenever mwait is available. As a result, since a vCPU in mwait
>> remains in guest context and does not yield to the hypervisor, an
>> offline vCPU *appears* to be 100% busy as viewed from ESXi, which
>> prevents the hypervisor from running other vCPUs or workloads on the
>> corresponding pCPU (particularly when vCPU - pCPU mappings are
>> statically defined by the user).
> 
> I would hope vCPU pinning is a mandatory thing when MWAIT passthrough it
> set?
> 
>> [ Note that such a vCPU is not
>> actually busy spinning though; it remains in mwait idle state in the
>> guest ].
>>
>> Fix this by overriding the CPU offline play_dead() callback for VMware
>> hypervisor, by putting the CPU in halt state (which actually yields to
>> the hypervisor), even if mwait support is available.
>>
>> Signed-off-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
>> ---
> 
>> +static void vmware_play_dead(void)
>> +{
>> +	play_dead_common();
>> +	tboot_shutdown(TB_SHUTDOWN_WFS);
>> +
>> +	/*
>> +	 * Put the vCPU going offline in halt instead of mwait (even
>> +	 * if mwait support is available), to make sure that the
>> +	 * offline vCPU yields to the hypervisor (which may not happen
>> +	 * with mwait, for example, if the guest's VMX is configured
>> +	 * to retain the vCPU in guest context upon mwait).
>> +	 */
>> +	hlt_play_dead();
>> +}
>>   #endif
>>   
>>   static __init int activate_jump_labels(void)
>> @@ -349,6 +365,7 @@ static void __init vmware_paravirt_ops_setup(void)
>>   #ifdef CONFIG_SMP
>>   		smp_ops.smp_prepare_boot_cpu =
>>   			vmware_smp_prepare_boot_cpu;
>> +		smp_ops.play_dead = vmware_play_dead;
>>   		if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
>>   					      "x86/vmware:online",
>>   					      vmware_cpu_online,
> 
> No real objection here; but would not something like the below fix the
> problem more generally? I'm thinking MWAIT passthrough for *any*
> hypervisor doesn't want play_dead to use it.
> 
> diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
> index f24227bc3220..166cb3aaca8a 100644
> --- a/arch/x86/kernel/smpboot.c
> +++ b/arch/x86/kernel/smpboot.c
> @@ -1759,6 +1759,8 @@ static inline void mwait_play_dead(void)
>   		return;
>   	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
>   		return;
> +	if (this_cpu_has(X86_FEATURE_HYPERVISOR))
> +		return;
>   	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
>   		return;
>   

With my Xen hat on I agree with this approach.


Juergen

[-- Attachment #1.1.2: OpenPGP public key --]
[-- Type: application/pgp-keys, Size: 3149 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]