Re: [PATCH net-next 16/17] net: sched: conditionally take rtnl lock on rules update path

[Stefano Brivio <sbrivio@redhat.com>]

On Tue, 13 Nov 2018 13:25:52 +0000
Vlad Buslov <vladbu@mellanox.com> wrote:

> On Tue 13 Nov 2018 at 09:40, Stefano Brivio <sbrivio@redhat.com> wrote:
> > Hi Vlad,
> >
> > On Mon, 12 Nov 2018 09:55:45 +0200
> > Vlad Buslov <vladbu@mellanox.com> wrote:
> >  
> >> @@ -179,9 +179,25 @@ static void tcf_proto_destroy_work(struct work_struct *work)
> >>  	rtnl_unlock();
> >>  }
> >>  
> >> +/* Helper function to lock rtnl mutex when specified condition is true and mutex
> >> + * hasn't been locked yet. Will set rtnl_held to 'true' before taking rtnl lock.
> >> + * Note that this function does nothing if rtnl is already held. This is
> >> + * intended to be used by cls API rules update API when multiple conditions
> >> + * could require rtnl lock and its state needs to be tracked to prevent trying
> >> + * to obtain lock multiple times.
> >> + */
> >> +
> >> +static void tcf_require_rtnl(bool cond, bool *rtnl_held)
> >> +{
> >> +	if (!*rtnl_held && cond) {
> >> +		*rtnl_held = true;
> >> +		rtnl_lock();
> >> +	}
> >> +}  
> >
> > I guess calls to this function are supposed to be serialised. If that's
> > the case (which is my tentative understanding so far), I would indicate
> > that in the comment.
> >
> > If that's not the case, you would be introducing a race I guess.
> >
> > Same applies to tcf_block_release() from 17/17.  
> 
> Hi Stefano,
> 
> Thank you for reviewing my code!
> 
> I did not intend for this function to be serialized. First argument to
> tcf_require_rtnl() is passed by value, and second argument is always a
> pointer to local stack-allocated value of the caller.

Yes, sorry, I haven't been terribly clear, that's what I meant by
serialised: it won't be called concurrently with the same *rtnl_held.

Perhaps the risk that somebody uses it that way is close to zero, so
I'm not even too sure this is worth a comment, but if you can come up
with a concise way of saying this, that would be nice.

> Same applies to tcf_block_release() - its arguments are Qdisc and block
> which support concurrency-safe reference counting, and pointer to local
> variable rtnl_held, which is not accessible to concurrent users.

Same there.

> What is the race in these cases? Am I missing something?

No, no race then. My only concern was:

thread A:                             thread B:
- x = false;
- tcf_require_rtnl(true, &x);         - tcf_require_rtnl(true, &x);
  - if (!*x && true)                    - if (!*x && true)
    - *x = true;
    - rtnl_lock()                         - *x = true;
                                          - rtnl_lock()

but this cannot happen as you explained.

-- 
Stefano