Re: [PATCH v2 3/6] x86/kvm: introduce per cpu vcpu masks

[Vitaly Kuznetsov]

Eduardo Habkost <ehabkost@xxxxxxxxxx> writes:

> On Fri, Sep 03, 2021 at 03:08:04PM +0200, Juergen Gross wrote:
>> In order to support high vcpu numbers per guest don't use on stack
>> vcpu bitmasks. As all those currently used bitmasks are not used in
>> functions subject to recursion it is fairly easy to replace them with
>> percpu bitmasks.
>> 
>> Disable preemption while such a bitmask is being used in order to
>> avoid double usage in case we'd switch cpus.
>> 
>> Signed-off-by: Juergen Gross <jgross@xxxxxxxx>
>> ---
>> V2:
>> - use local_lock() instead of preempt_disable() (Paolo Bonzini)
>> ---
>>  arch/x86/include/asm/kvm_host.h | 10 ++++++++++
>>  arch/x86/kvm/hyperv.c           | 25 ++++++++++++++++++-------
>>  arch/x86/kvm/irq_comm.c         |  9 +++++++--
>>  arch/x86/kvm/x86.c              | 22 +++++++++++++++++++++-
>>  4 files changed, 56 insertions(+), 10 deletions(-)
>> 
>> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
>> index 3513edee8e22..a809a9e4fa5c 100644
>> --- a/arch/x86/include/asm/kvm_host.h
>> +++ b/arch/x86/include/asm/kvm_host.h
>> @@ -15,6 +15,7 @@
>>  #include <linux/cpumask.h>
>>  #include <linux/irq_work.h>
>>  #include <linux/irq.h>
>> +#include <linux/local_lock.h>
>>  
>>  #include <linux/kvm.h>
>>  #include <linux/kvm_para.h>
>> @@ -1591,6 +1592,15 @@ extern bool kvm_has_bus_lock_exit;
>>  /* maximum vcpu-id */
>>  unsigned int kvm_max_vcpu_id(void);
>>  
>> +/* per cpu vcpu bitmasks, protected by kvm_pcpu_mask_lock */
>> +DECLARE_PER_CPU(local_lock_t, kvm_pcpu_mask_lock);
>> +extern unsigned long __percpu *kvm_pcpu_vcpu_mask;
>> +#define KVM_VCPU_MASK_SZ	\
>> +	(sizeof(*kvm_pcpu_vcpu_mask) * BITS_TO_LONGS(KVM_MAX_VCPUS))
>> +extern u64 __percpu *kvm_hv_vp_bitmap;
>> +#define KVM_HV_MAX_SPARSE_VCPU_SET_BITS DIV_ROUND_UP(KVM_MAX_VCPUS, 64)
>> +#define KVM_HV_VPMAP_SZ		(sizeof(u64) * KVM_HV_MAX_SPARSE_VCPU_SET_BITS)
>
> I have just realized that the Hyper-V sparse bitmap format can
> support only up to 4096 CPUs, and the current implementation of
> sparse_set_to_vcpu_mask() won't even work correctly if
> KVM_MAX_VCPUS is larger than 4096.

Nice catch! Indeed, we can only encode 64 'banks' with 64 vCPUs each. We
need to enforce the limit somehow. As a big hammer, I can suggest to
fail kvm_hv_vcpu_init() and write to HV_X64_MSR_VP_INDEX for vCPUs above
4095 for now (I seriously doubt there's real need to run such big
Windows guests anywhere but who knows).

>
> This means vp_bitmap can't and will never be larger than 512
> bytes.  Isn't a per-CPU variable for vp_bitmap overkill in this
> case?

Yes, it's OK to allocate 512 bytes on stack.

-- 
Vitaly