Re: Fwd: a8d5875ce5 ("Default enable RCU list lockdep debugging with .."): WARNING: suspicious RCU usage

[Madhuparna Bhowmik <madhuparnabhowmik10@gmail.com>]

On Wed, Apr 29, 2020 at 03:40:58PM -0700, Paul E. McKenney wrote:
> On Wed, Apr 29, 2020 at 03:34:33PM +0530, Madhuparna Bhowmik wrote:
> > On Tue, Apr 28, 2020 at 08:59:01AM -0400, Mimi Zohar wrote:
> > > On Tue, 2020-04-28 at 16:53 +0530, Madhuparna Bhowmik wrote:
> > > > On Mon, Apr 27, 2020 at 08:58:26PM -0400, Mimi Zohar wrote:
> > > > > [Cc'ing Matthew Garrett)
> > > > > 
> > > > > Hi Madhuparna,
> > > > > 
> > > > > On Sat, 2020-04-25 at 16:33 +0530, Madhuparna Bhowmik wrote:
> > > > > > Hi,
> > > > > > 
> > > > > > This is regarding the warning reported by kernel test bot regarding
> > > > > > suspicious RCU usage.
> > > > > > Using a simple git grep, I can only see the following usage of RCU:
> > > > > > 
> > > > > > evm_crypto.c:   list_for_each_entry_rcu(xattr, &evm_config_xattrnames,
> > > > > > list) {
> > > > > > evm_main.c:     list_for_each_entry_rcu(xattr, &evm_config_xattrnames,
> > > > > > list) {
> > > > > > evm_main.c:     list_for_each_entry_rcu(xattr, &evm_config_xattrnames,
> > > > > > list) {
> > > > > > evm_secfs.c:    list_add_tail_rcu(&xattr->list, &evm_config_xattrnames);
> > > > > > 
> > > > > > So, the evm_config_xattrnames list is traversed using
> > > > > > list_for_each_entry_rcu() but without the protection of rcu_read_lock()?
> > > > > > If these are not really RCU read-side CS, and other locks are held then
> > > > > > there is no need to use list_for_each_entry_rcu().
> > > > > > And maybe we can completely remove the usage of rcu primitives here.
> > > > > > Or if there is a bug and rcu_read_lock() should be held, please let me know
> > > > > > and I can try fixing this.
> > > > > 
> > > > > Thank you for forwarding this report.  The list of EVM xattrs is
> > > > > protected by the xattr_list_mutex, which is used when reading or
> > > > > appending to the EVM list itself.  Entries in the list can not be
> > > > > removed.
> > > > >
> > > > Hi Mimi,
> > > > 
> > > > Thank you for your reply.
> > > > So, if the list is protected by xattr_list mutex and it is used during
> > > > both reading and writing to the list, can we remove the usage of RCU
> > > > here? 
> > > 
> > > I should have said the mutex is used when cat'ing the securityfs file
> > > (security/integrity/evm/evm_xattrs) and when adding to the list, but
> > > not in the above cases when walking the list.
> > >
> > > > Since the read side critical section is already protected by the
> > > > xattr_list mutex, we do not need list_for_each_entry_rcu() to read the
> > > > list. Also, we can just simply add to the list using list_add_tail(),
> > > > RCU primitives are not really required here.
> > > > 
> > > > Please let me know is this is fine, and I can send a patch removing the
> > > > usage of RCU here.
> > > 
> > > Matthew, please correct me if I'm wrong, the reason it is safe, is not
> > > because there is a mutex, but because entries are never removed from
> > > the list.
> > >
> > Alright, I understood the case here. So entries are only added to the
> > tail of the list and never deleted. And that's why it is safe for
> > readers and writers to execute concurrently even without the mutex.
> > 
> > However, RCU would still complain if no lock or rcu_read_lock is not
> > held.
> > 
> > Should I cc Paul McKenney about this case, he is the RCU Maintainer and
> > usually replies pretty fast.
> > He would be able to correctly suggest how to fix the RCU usage here.
> > 
> > Let me know if this is okay.
> 
> Apparently, it is not necessary to CC me.  ;-)
>
Surprised!

> You do of course need the code to use the RCU variants of list_add*().
> And also list_for_each_entry_rcu(), as in the current code.
> 
> There are several options, none of them perfect:
> 
> 1.	Add (not otherwise needed) calls to rcu_read_lock() and
> 	rcu_read_unlock() and leave list_for_each_entry_rcu() unchanged.
> 
> 2.	Add "true" for the optional fourth argument to
> 	list_for_each_entry_rcu().  This will suppress the complaints,
> 	but would (incorrectly) continue to do so should this code change
> 	so as to be able to delete form this list.
> 
> 3.	Switch from list_for_each_entry_rcu() to its lockless counterpart,
> 	list_for_each_entry_lockless().  This is simiar to #2 above, but
> 	at least the name lets people know that something unusual is up.
> 
> If it was my code, I would take door #3.  ;-)
>
Thanks a lot for your inputs on this. I will send a patch soon.

Regards,
Madhuparna
> 							Thanx, Paul