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

[Paolo Bonzini]

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.

Paul, do you have any objections to the idea?  Thanks,

Paolo