Circular lockdep in kvm_reset_vcpu() ?

[Cristian Marussi]

On Mon, Mar 13, 2023 at 02:09:55PM +0000, Marc Zyngier wrote:
> On 2023-03-13 10:09, Cristian Marussi wrote:
> > On Sat, Feb 11, 2023 at 12:56:41AM +0000, Oliver Upton wrote:
> > > Hi Jeremy,
> > > 
> > 
> > Hi,
> > 
> > > On Fri, Feb 10, 2023 at 11:46:36AM -0600, Jeremy Linton wrote:
> > > > Hi,
> > > >
> > > > I saw this pop yesterday:
> > > 
> > > You and me both actually! Shame on me, I spoke off-list about this
> > > with
> > > Marc in passing. Thanks for sending along the report.
> > > 
> > > > [   78.333360] ======================================================
> > > > [   78.339541] WARNING: possible circular locking dependency detected
> > > > [   78.345721] 6.2.0-rc7+ #19 Not tainted
> > > > [   78.349470] ------------------------------------------------------
> > > > [   78.355647] qemu-system-aar/859 is trying to acquire lock:
> > > > [   78.361130] ffff5aa69269eba0 (&host_kvm->lock){+.+.}-{3:3}, at:
> > > > kvm_reset_vcpu+0x34/0x274
> > > > [   78.369344]
> > > > [   78.369344] but task is already holding lock:
> > > > [   78.375182] ffff5aa68768c0b8 (&vcpu->mutex){+.+.}-{3:3}, at:
> > > > kvm_vcpu_ioctl+0x8c/0xba0
> > > 
> > > [...]
> > > 
> > > > It appears to be triggered by the new commit 42a90008f890a ('KVM: Ensure
> > > > lockdep knows about kvm->lock vs. vcpu->mutex ordering rule') which is
> > > > detecting the vcpu lock grabbed by kvm_vcpu_ioctl() and then the kvm mutext
> > > > grabbed by kvm_reset_vcpu().
> > > 
> > > Right, this commit gave lockdep what it needed to smack us on the head
> > > for getting the locking wrong in the arm64 side.
> > > 
> > > As gross as it might be, the right direction is likely to have our own
> > > lock in kvm_arch that we can acquire while holding the vcpu mutex.
> > > I'll
> > > throw a patch at the list once I get done testing it.
> > > 
> > 
> > I just hit this using a v6.3-rc2 and a mainline kvmtool.
> > 
> > In my case, though, the guest does not even boot if I use more than 1
> > vcpu, which
> > I suppose triggers effectively the reported possible deadlock, i.e.:
> > 
> > root/lkvm_master run -c 4 -m 4096 -k /root/Image_guest -d
> > /root/disk_debian_buster_guest.img -p "loglevel=8"
> >   # lkvm run -k /root/Image_guest -m 4096 -c 4 --name guest-288
> > ....<HANGS FOREVER>
> 
> Pass earlycon to the guest for a start.
> 
> I seriously doubt someone has actually seen a deadlock, because
> the issue has been there for at least the past 7 years...
> 
> And -rc2 works just fine here.

Thanks, I'll dig deeper what's going on un my setup.
Cristian