Re: [PATCH v2] KVM: Synchronize accesses to the kvm->devices list

[Christoffer Dall <christoffer.dall@linaro.org>]

On Wed, Aug 03, 2016 at 01:24:41PM +0200, Paolo Bonzini wrote:
> 
> 
> On 03/08/2016 12:39, Christoffer Dall wrote:
> > The list of KVM devices is not currently protected with any lock.  As
> > far as I can see nothing prevents multiple threads from creating devices
> > on the KVM fd simultaenously, potentially leading to corruption.
> > Further, we already have VFIO looping over this list without any form of
> > synchronization, and we are about to add similar functionality for the
> > ITS on arm64.
> > 
> > Note that the traversal in kvm_destroy_devices does not appear to be a
> > problem because only the last reference to the KVM struct can actually
> > perform this action.
> > 
> > I use a mutex to protect the list, since the arm64 ITS code is going to
> > need to iterate over the devices (to find all ITSes) and register KVM IO
> > devices for each one, which in turn grabs another mutex and kmallocs
> > stuff.
> > 
> > Signed-off-by: Christoffer Dall <christoffer.dall@linaro.org>
> > ---
> > Changes since v1:
> >   - My RCU knowledge was too rusty and was both incorrectly implemented
> >     and would not address the arm64 need.  Just use a mutex instead.
> > 
> >  include/linux/kvm_host.h |  1 +
> >  virt/kvm/kvm_main.c      |  8 ++++++++
> >  virt/kvm/vfio.c          | 11 +++++++++--
> >  3 files changed, 18 insertions(+), 2 deletions(-)
> > 
> > diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> > index aafd702..de32196 100644
> > --- a/include/linux/kvm_host.h
> > +++ b/include/linux/kvm_host.h
> > @@ -421,6 +421,7 @@ struct kvm {
> >  	long mmu_notifier_count;
> >  #endif
> >  	long tlbs_dirty;
> > +	struct mutex devices_lock; /* protects the devices list */
> >  	struct list_head devices;
> >  	struct dentry *debugfs_dentry;
> >  	struct kvm_stat_data **debugfs_stat_data;
> > diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> > index 61b31a5..33e0579 100644
> > --- a/virt/kvm/kvm_main.c
> > +++ b/virt/kvm/kvm_main.c
> > @@ -616,6 +616,7 @@ static struct kvm *kvm_create_vm(unsigned long type)
> >  	mutex_init(&kvm->irq_lock);
> >  	mutex_init(&kvm->slots_lock);
> >  	atomic_set(&kvm->users_count, 1);
> > +	mutex_init(&kvm->devices_lock);
> >  	INIT_LIST_HEAD(&kvm->devices);
> >  
> >  	r = kvm_arch_init_vm(kvm, type);
> > @@ -694,6 +695,11 @@ static void kvm_destroy_devices(struct kvm *kvm)
> >  {
> >  	struct kvm_device *dev, *tmp;
> >  
> > +	/*
> > +	 * We do not need to take the devices_lock here, because nobody else
> > +	 * has a reference to the struct kvm at this point and therefore
> > +	 * cannot access the devices list anyhow.
> > +	 */
> >  	list_for_each_entry_safe(dev, tmp, &kvm->devices, vm_node) {
> >  		list_del(&dev->vm_node);
> >  		dev->ops->destroy(dev);
> > @@ -2842,7 +2848,9 @@ static int kvm_ioctl_create_device(struct kvm *kvm,
> >  		return ret;
> >  	}
> >  
> > +	mutex_lock(&kvm->devices_lock);
> >  	list_add(&dev->vm_node, &kvm->devices);
> > +	mutex_unlock(&kvm->devices_lock);
> 
> This is not enough for the VFIO test to be robust.  I think it's easier
> to move ops->create entirely under the kvm_lock.  This requires some
> auditing of the create functions, but many of them (e.g.
> kvm_vgic_create) already use that lock.
> 
> The only complex case is xics.  It takes kvm_lock only for part of the
> initialization, and the call to xics_debugfs_init would be better moved
> outside kvm_lock.  For that we can add an ops->init(dev) function that
> is called outside kvm_lock and cannot fail.
> 
Why do we need to move xics_debugfs_init outside of kvm->lock()? Merely
because it could sleep forever and we don't want that to block
everything or does it loop around and deadlock somehow that I cannot
easily see?

Thanks,
-Christoffer