Commit de5ed007 authored by Artemy Kovalyov's avatar Artemy Kovalyov Committed by Jason Gunthorpe
Browse files

IB/mlx5: Fix implicit ODP race 

Following race may occur because of the call_srcu and the placement of
the synchronize_srcu vs the xa_erase.

CPU0				   CPU1

mlx5_ib_free_implicit_mr:	   destroy_unused_implicit_child_mr:
 xa_erase(odp_mkeys)
 synchronize_srcu()
				    xa_lock(implicit_children)
				    if (still in xarray)
				       atomic_inc()
				       call_srcu()
				    xa_unlock(implicit_children)
 xa_erase(implicit_children):
   xa_lock(implicit_children)
   __xa_erase()
   xa_unlock(implicit_children)

 flush_workqueue()
				   [..]
				    free_implicit_child_mr_rcu:
				     (via call_srcu)
				      queue_work()

 WARN_ON(atomic_read())
				   [..]
				    free_implicit_child_mr_work:
				     (via wq)
				      free_implicit_child_mr()
 mlx5_mr_cache_invalidate()
				     mlx5_ib_update_xlt() <-- UMR QP fail
				     atomic_dec()

The wait_event() solves the race because it blocks until
free_implicit_child_mr_work() completes.

Fixes: 5256edcb ("RDMA/mlx5: Rework implicit ODP destroy")
Link: https://lore.kernel.org/r/20200227113918.94432-1-leon@kernel.org


Signed-off-by: default avatarArtemy Kovalyov <artemyko@mellanox.com>
Reviewed-by: default avatarJason Gunthorpe <jgg@mellanox.com>
Signed-off-by: default avatarLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: default avatarJason Gunthorpe <jgg@mellanox.com>
parent 817a68a6

drivers/infiniband/hw/mlx5/mlx5_ib.h

+1 −0
Original line number Diff line number Diff line
@@ -636,6 +636,7 @@ struct mlx5_ib_mr { 


	/* For ODP and implicit */

	atomic_t		num_deferred_work;

	wait_queue_head_t       q_deferred_work;

	struct xarray		implicit_children;

	union {

		struct rcu_head rcu;

drivers/infiniband/hw/mlx5/odp.c

+7 −10
Original line number Diff line number Diff line
@@ -235,7 +235,8 @@ static void free_implicit_child_mr(struct mlx5_ib_mr *mr, bool need_imr_xlt) 
	mr->parent = NULL;

	mlx5_mr_cache_free(mr->dev, mr);

	ib_umem_odp_release(odp);

	atomic_dec(&imr->num_deferred_work);

	if (atomic_dec_and_test(&imr->num_deferred_work))

		wake_up(&imr->q_deferred_work);

}



static void free_implicit_child_mr_work(struct work_struct *work)
@@ -554,6 +555,7 @@ struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd, 
	imr->umem = &umem_odp->umem;

	imr->is_odp_implicit = true;

	atomic_set(&imr->num_deferred_work, 0);

	init_waitqueue_head(&imr->q_deferred_work);

	xa_init(&imr->implicit_children);



	err = mlx5_ib_update_xlt(imr, 0,
@@ -611,10 +613,7 @@ void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr) 
	 * under xa_lock while the child is in the xarray. Thus at this point

	 * it is only decreasing, and all work holding it is now on the wq.

	 */

	if (atomic_read(&imr->num_deferred_work)) {

		flush_workqueue(system_unbound_wq);

		WARN_ON(atomic_read(&imr->num_deferred_work));

	}

	wait_event(imr->q_deferred_work, !atomic_read(&imr->num_deferred_work));



	/*

	 * Fence the imr before we destroy the children. This allows us to
@@ -645,10 +644,7 @@ void mlx5_ib_fence_odp_mr(struct mlx5_ib_mr *mr) 
	/* Wait for all running page-fault handlers to finish. */

	synchronize_srcu(&mr->dev->odp_srcu);



	if (atomic_read(&mr->num_deferred_work)) {

		flush_workqueue(system_unbound_wq);

		WARN_ON(atomic_read(&mr->num_deferred_work));

	}

	wait_event(mr->q_deferred_work, !atomic_read(&mr->num_deferred_work));



	dma_fence_odp_mr(mr);

}
@@ -1720,7 +1716,8 @@ static void destroy_prefetch_work(struct prefetch_mr_work *work) 
	u32 i;



	for (i = 0; i < work->num_sge; ++i)

		atomic_dec(&work->frags[i].mr->num_deferred_work);

		if (atomic_dec_and_test(&work->frags[i].mr->num_deferred_work))

			wake_up(&work->frags[i].mr->q_deferred_work);

	kvfree(work);

}

CRA Git | Maintained and supported by SUSTech CRA and CCSE