[Linux-kernel-mentees] [PATCH v2] kvm: Fix false positive RCU usage warning

[Joel Fernandes]

On Tue, Jun 23, 2020 at 11:19:20PM +0530, Madhuparna Bhowmik wrote:
> On Tue, Jun 23, 2020 at 08:39:01AM -0700, Paul E. McKenney wrote:
> > On Tue, Jun 23, 2020 at 09:00:36PM +0530, Madhuparna Bhowmik wrote:
> > > On Tue, Jun 23, 2020 at 11:02:36AM -0400, Joel Fernandes wrote:
> > > > On Tue, Jun 23, 2020 at 09:39:53AM +0200, Paolo Bonzini wrote:
> > > > > On 16/05/20 10:22, madhuparnabhowmik10 at gmail.com wrote:
> > > > > > From: Madhuparna Bhowmik <madhuparnabhowmik10 at gmail.com>
> > > > > > 
> > > > > > Fix the following false positive warnings:
> > > > > > 
> > > > > > [ 9403.765413][T61744] =============================
> > > > > > [ 9403.786541][T61744] WARNING: suspicious RCU usage
> > > > > > [ 9403.807865][T61744] 5.7.0-rc1-next-20200417 #4 Tainted: G             L
> > > > > > [ 9403.838945][T61744] -----------------------------
> > > > > > [ 9403.860099][T61744] arch/x86/kvm/mmu/page_track.c:257 RCU-list traversed in non-reader section!!
> > > > > > 
> > > > > > and
> > > > > > 
> > > > > > [ 9405.859252][T61751] =============================
> > > > > > [ 9405.859258][T61751] WARNING: suspicious RCU usage
> > > > > > [ 9405.880867][T61755] -----------------------------
> > > > > > [ 9405.911936][T61751] 5.7.0-rc1-next-20200417 #4 Tainted: G             L
> > > > > > [ 9405.911942][T61751] -----------------------------
> > > > > > [ 9405.911950][T61751] arch/x86/kvm/mmu/page_track.c:232 RCU-list traversed in non-reader section!!
> > > > > > 
> > > > > > Since srcu read lock is held, these are false positive warnings.
> > > > > > Therefore, pass condition srcu_read_lock_held() to
> > > > > > list_for_each_entry_rcu().
> > > > > > 
> > > > > > Reported-by: kernel test robot <lkp at intel.com>
> > > > > > Signed-off-by: Madhuparna Bhowmik <madhuparnabhowmik10 at gmail.com>
> > > > > > ---
> > > > > > v2:
> > > > > > -Rebase v5.7-rc5
> > > > > > 
> > > > > >  arch/x86/kvm/mmu/page_track.c | 6 ++++--
> > > > > >  1 file changed, 4 insertions(+), 2 deletions(-)
> > > > > > 
> > > > > > diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
> > > > > > index ddc1ec3bdacd..1ad79c7aa05b 100644
> > > > > > --- a/arch/x86/kvm/mmu/page_track.c
> > > > > > +++ b/arch/x86/kvm/mmu/page_track.c
> > > > > > @@ -229,7 +229,8 @@ void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
> > > > > >  		return;
> > > > > >  
> > > > > >  	idx = srcu_read_lock(&head->track_srcu);
> > > > > > -	hlist_for_each_entry_rcu(n, &head->track_notifier_list, node)
> > > > > > +	hlist_for_each_entry_rcu(n, &head->track_notifier_list, node,
> > > > > > +				srcu_read_lock_held(&head->track_srcu))
> > > > > >  		if (n->track_write)
> > > > > >  			n->track_write(vcpu, gpa, new, bytes, n);
> > > > > >  	srcu_read_unlock(&head->track_srcu, idx);
> > > > > > @@ -254,7 +255,8 @@ void kvm_page_track_flush_slot(struct kvm *kvm, struct kvm_memory_slot *slot)
> > > > > >  		return;
> > > > > >  
> > > > > >  	idx = srcu_read_lock(&head->track_srcu);
> > > > > > -	hlist_for_each_entry_rcu(n, &head->track_notifier_list, node)
> > > > > > +	hlist_for_each_entry_rcu(n, &head->track_notifier_list, node,
> > > > > > +				srcu_read_lock_held(&head->track_srcu))
> > > > > >  		if (n->track_flush_slot)
> > > > > >  			n->track_flush_slot(kvm, slot, n);
> > > > > >  	srcu_read_unlock(&head->track_srcu, idx);
> > > > > > 
> > > > > 
> > > > > Hi, sorry for the delay in reviewing this patch.  I would like to ask
> > > > > Paul about it.
> > > > > 
> > > > > While you're correctly fixing a false positive, hlist_for_each_entry_rcu
> > > > > would have a false _negative_ if you called it under
> > > > > rcu_read_lock/unlock and the data structure was protected by SRCU.  This
> > > > > is why for example srcu_dereference is used instead of
> > > > > rcu_dereference_check, and why srcu_dereference uses
> > > > > __rcu_dereference_check (with the two underscores) instead of
> > > > > rcu_dereference_check.  Using rcu_dereference_check would add an "||
> > > > > rcu_read_lock_held()" to the condition which is wrong.
> > > > > 
> > > > > I think instead you should add hlist_for_each_srcu and
> > > > > hlist_for_each_entry_srcu macro to include/linux/rculist.h.
> > > > > 
> > > > > There is no need for equivalents of hlist_for_each_entry_continue_rcu
> > > > > and hlist_for_each_entry_from_rcu, because they use rcu_dereference_raw.
> > > > >  However, it's not documented why they do so.
> > > > 
> > > > You are right, this patch is wrong, we need a new SRCU list macro to do the
> > > > right thing which would also get rid of the last list argument.
> > > >
> > > Can we really get rid of the last argument? We would need the
> > > srcu_struct right for checking?
> > 
> > Agreed!  However, the API could be simplified by passing in a pointer to
> > the srcu_struct instead of a lockdep expression.  An optional lockdep
> > expression might still be helpful for calls from the update side,
> > of course.
> >
> Sure, I will work on this.

Cool, thanks !!!

 - Joel