Re: [PATCH v3] KVM: remove buggy vcpu id check on vcpu creation

[Greg Kurz]

On Thu, 21 Apr 2016 14:26:19 +0200
Cornelia Huck <cornelia.huck@xxxxxxxxxx> wrote:

> On Thu, 21 Apr 2016 13:29:58 +0200
> Greg Kurz <gkurz@xxxxxxxxxxxxxxxxxx> wrote:
> 
> > On Wed, 20 Apr 2016 20:29:09 +0200
> > Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx> wrote:
> >   
> > > 2016-04-20 17:44+0200, Greg Kurz:  
> > > > Commit 338c7dbadd26 ("KVM: Improve create VCPU parameter (CVE-2013-4587)")
> > > > introduced a check to prevent potential kernel memory corruption in case
> > > > the vcpu id is too great.
> > > > 
> > > > Unfortunately this check assumes vcpu ids grow in sequence with a common
> > > > difference of 1, which is wrong: archs are free to use vcpu id as they fit.
> > > > For example, QEMU originated vcpu ids for PowerPC cpus running in boot3s_hv
> > > > mode, can grow with a common difference of 2, 4 or 8: if KVM_MAX_VCPUS is
> > > > 1024, guests may be limited down to 128 vcpus on POWER8.
> > > > 
> > > > This means the check does not belong here and should be moved to some arch
> > > > specific function: kvm_arch_vcpu_create() looks like a good candidate.
> > > > 
> > > > ARM and s390 already have such a check.
> > > > 
> > > > I could not spot any path in the PowerPC or common KVM code where a vcpu
> > > > id is used as described in the above commit: I believe PowerPC can live
> > > > without this check.    
> > > 
> > > The only problematic path I see is kvm_get_vcpu_by_id(), which returns
> > > NULL for any id above KVM_MAX_VCPUS.  
> > 
> > Oops my bad, I started to work on a 4.4 tree and I missed this check brought
> > by commit c896939f7cff (KVM: use heuristic for fast VCPU lookup by id).
> > 
> > But again, I believe the check is wrong there also: the changelog just mentions
> > this is a fastpath for the usual case where "VCPU ids match the array index"...
> > why does the patch add a NULL return path if id >= KVM_MAX_VCPUS ?  
> 
> Probably because noone considered power :)
> 

No surprise but the return path is a bit overkill anyway :)

> >   
> > > kvm_vm_ioctl_create_vcpu() uses kvm_get_vcpu_by_id() to check for
> > > duplicate ids, so PowerPC could end up with many VCPUs of the same id.
> > > I'm not sure what could fail, but code doesn't expect this situation.
> > > Patching kvm_get_vcpu_by_id() is easy, though.
> > >   
> > 
> > Something like this ?
> > 
> > 	if (id < 0)
> > 		return NULL;
> > 	if (id < KVM_MAX_VCPUS)
> > 		vcpu = kvm_get_vcpu(kvm, id);
> > 
> > In the same patch ?
> >   
> > > Second issue is that Documentation/virtual/kvm/api.txt says
> > >   4.7 KVM_CREATE_VCPU
> > >   [...]
> > >   This API adds a vcpu to a virtual machine.  The vcpu id is a small
> > >   integer in the range [0, max_vcpus).
> > >   
> > 
> > Yeah and I find the meaning of max_vcpus is unclear.
> > 
> > Here it is considered as a limit for vcpu id, but if you look at the code,
> > KVM_MAX_VCPUS is also used as a limit for the number of vcpus:
> > 
> > virt/kvm/kvm_main.c:    if (atomic_read(&kvm->online_vcpus) == KVM_MAX_VCPUS) {
> >   
> > > so we'd remove those two lines and change the API too.  The change would
> > > be somewhat backward compatible, but doesn't PowerPC use high vcpu_id
> > > just because KVM is lacking an API to set DT ID?  
> > 
> > This is related to a limitation when running in book3s_hv mode with cpus
> > that support SMT (multiple HW threads): all HW threads within a core
> > cannot be running in different guests at the same time. 
> > 
> > We solve this by using a vcpu numbering scheme as follows:
> > 
> > vcpu_id[N] = (N * thread_per_core_guest) / threads_per_core_host + N % threads_per_core_guest
> > 
> > where N means "the Nth vcpu presented to the guest". This allows to have groups of vcpus
> > that can be scheduled to run on the same real core.
> > 
> > So, in the "worst" case where we want to run a guest with 1 thread/core and the host
> > has 8 threads/core, we will need the vcpu_id limit to be 8*KVM_MAX_VCPUS.
> >   
> > > x86 (APIC ID) is affected by this and ARM (MP ID) probably too.
> > >   
> > 
> > x86 is limited to KVM_MAX_VCPUS (== 255) vcpus: it won't be affected if we also
> > patch kvm_get_vcpu_by_id() like suggested above.
> > 
> > Depending on the platform, ARM can be limited to VGIC_V3_MAX_CPUS (== 255) or
> > VGIC_V8_MAX_CPUS (== 8). I guess it won't be affected either.  
> 
> For s390, it's either 64 (no esca) or 248 (esca).
> 

And it is CONFIG_NR_CPUS for powerpc, i.e. 2048 per default on powernv.

But the problem here is more: can we compare the number of vcpus with vcpu ids ?

> >   
> > > (Maybe it is time to decouple VCPU ID used in KVM interfaces from
> > >  architecture dependent CPU ID that the guest uses ...  
> > 
> > Maybe... I did not get that far.  
> 
> It seems that the various architectures are more different than I
> thought... wasn't aware of the complicated situation on power, for
> example.

Yeah, and I think moving these vcpu id checks to the archs allows to
solve the problem and confine the complexity to the powerpc code.