On Thu, 19 Jan 2012, Eric Dumazet wrote:
> Le jeudi 19 janvier 2012 à 04:28 -0800, Tejun Heo a écrit :
> > Hello,
> >
> > On Wed, Jan 18, 2012 at 11:33 PM, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> > > Interesting, but should be a patch on its own.
> >
> > Yeap, agreed.
Okay, in that case I'd better split into three (idr, revert, remove lock).
I'll send those three in a moment. I've also slipped an RCU comment from
idr_find into idr_get_next, and put the Acks in all three.
> >
> > > Maybe other idr users can benefit from your idea as well, if patch is
> > > labeled "idr: allow idr_get_next() from rcu_read_lock" or something...
> > >
> > > I suggest introducing idr_get_next_rcu() helper to make the check about
> > > rcu cleaner.
> > >
> > > idr_get_next_rcu(...)
> > > {
> > > WARN_ON_ONCE(!rcu_read_lock_held());
> > > return idr_get_next(...);
> > > }
> >
> > Hmmm... I don't know. Does having a separate set of interface help
> > much? It's easy to avoid/miss the test by using the other one. If we
> > really worry about it, maybe indicating which locking is to be used
> > during init is better? We can remember the lockdep map and trigger
> > WARN_ON_ONCE() if neither the lock or RCU read lock is held.
>
>
> There is a rcu_dereference_raw(ptr) in idr_get_next()
>
> This could be changed to rcu_dereference_check(ptr, condition) to get
> lockdep support for free :)
>
> [ condition would be the appropriate
> lockdep_is_held(&the_lock_protecting_my_idr) or 'I use the rcu variant'
> and I hold rcu_read_lock ]
>
> This would need to add a 'condition' parameter to idr_gen_next(), but we
> have very few users in kernel at this moment.
idr_get_next() was introduced for memcg, and has only one other user
user in the tree (drivers/mtd/mtdcore.c, which uses a mutex to lock it).
With the RCU fix, idr_get_next() becomes very much like idr_find().
I'll leave any fiddling with their interfaces to you guys.
Hugh