On 2/11/22 13:00, Jason Gunthorpe wrote: > On Tue, Feb 08, 2022 at 03:16:35PM -0600, Bob Pearson wrote: > >> @@ -62,7 +61,7 @@ static int rxe_mcast_get_grp(struct rxe_dev *rxe, union ib_gid *mgid, >> if (rxe->attr.max_mcast_qp_attach == 0) >> return -EINVAL; >> >> - write_lock_bh(&pool->pool_lock); >> + spin_lock_bh(&rxe->mcg_lock); > >> grp = rxe_pool_get_key_locked(pool, mgid); > > Now this calls rxe_pool_get_key_locked() without the lock :\ > > This is all fixed up a few patches later, so it only hurts > bisect-ability, do you want to fix it? > > I looked up to 'Remove mcg from rxe pools' and it seems fine to me > otherwise > > Thanks, > Jason It is locked. Just a different lock. All uses of the red-black tree in rxe_pool for this object pool will be using the same lock. And as you say we are heading towards replacing it all. Just leave it. Bob
