[PATCH 3.2 114/199] net/mlx4_core: Fix racy CQ (Completion Queue) free

[Ben Hutchings]

3.2.87-rc1 review patch.  If anyone has any objections, please let me know.

------------------

From: Jack Morgenstein <jackm@xxxxxxxxxxxxxxxxxx>

commit 291c566a28910614ce42d0ffe82196eddd6346f4 upstream.

In function mlx4_cq_completion() and mlx4_cq_event(), the
radix_tree_lookup requires a rcu_read_lock.
This is mandatory: if another core frees the CQ, it could
run the radix_tree_node_rcu_free() call_rcu() callback while
its being used by the radix tree lookup function.

Additionally, in function mlx4_cq_event(), since we are adding
the rcu lock around the radix-tree lookup, we no longer need to take
the spinlock. Also, the synchronize_irq() call for the async event
eliminates the need for incrementing the cq reference count in
mlx4_cq_event().

Other changes:
1. In function mlx4_cq_free(), replace spin_lock_irq with spin_lock:
   we no longer take this spinlock in the interrupt context.
   The spinlock here, therefore, simply protects against different
   threads simultaneously invoking mlx4_cq_free() for different cq's.

2. In function mlx4_cq_free(), we move the radix tree delete to before
   the synchronize_irq() calls. This guarantees that we will not
   access this cq during any subsequent interrupts, and therefore can
   safely free the CQ after the synchronize_irq calls. The rcu_read_lock
   in the interrupt handlers only needs to protect against corrupting the
   radix tree; the interrupt handlers may access the cq outside the
   rcu_read_lock due to the synchronize_irq calls which protect against
   premature freeing of the cq.

3. In function mlx4_cq_event(), we change the mlx_warn message to mlx4_dbg.

4. We leave the cq reference count mechanism in place, because it is
   still needed for the cq completion tasklet mechanism.

Fixes: 6d90aa5cf17b ("net/mlx4_core: Make sure there are no pending async events when freeing CQ")
Fixes: 225c7b1feef1 ("IB/mlx4: Add a driver Mellanox ConnectX InfiniBand adapters")
Signed-off-by: Jack Morgenstein <jackm@xxxxxxxxxxxxxxxxxx>
Signed-off-by: Matan Barak <matanb@xxxxxxxxxxxx>
Signed-off-by: Tariq Toukan <tariqt@xxxxxxxxxxxx>
Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
[bwh: Backported to 3.2: adjust context]
Signed-off-by: Ben Hutchings <ben@xxxxxxxxxxxxxxx>
---
 drivers/net/ethernet/mellanox/mlx4/cq.c | 38 +++++++++++++++++----------------
 1 file changed, 20 insertions(+), 18 deletions(-)

--- a/drivers/net/ethernet/mellanox/mlx4/cq.c
+++ b/drivers/net/ethernet/mellanox/mlx4/cq.c
@@ -78,13 +78,19 @@ void mlx4_cq_completion(struct mlx4_dev
 {
 	struct mlx4_cq *cq;
 
+	rcu_read_lock();
 	cq = radix_tree_lookup(&mlx4_priv(dev)->cq_table.tree,
 			       cqn & (dev->caps.num_cqs - 1));
+	rcu_read_unlock();
+
 	if (!cq) {
 		mlx4_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
 		return;
 	}
 
+	/* Acessing the CQ outside of rcu_read_lock is safe, because
+	 * the CQ is freed only after interrupt handling is completed.
+	 */
 	++cq->arm_sn;
 
 	cq->comp(cq);
@@ -95,23 +101,19 @@ void mlx4_cq_event(struct mlx4_dev *dev,
 	struct mlx4_cq_table *cq_table = &mlx4_priv(dev)->cq_table;
 	struct mlx4_cq *cq;
 
-	spin_lock(&cq_table->lock);
-
+	rcu_read_lock();
 	cq = radix_tree_lookup(&cq_table->tree, cqn & (dev->caps.num_cqs - 1));
-	if (cq)
-		atomic_inc(&cq->refcount);
-
-	spin_unlock(&cq_table->lock);
+	rcu_read_unlock();
 
 	if (!cq) {
-		mlx4_warn(dev, "Async event for bogus CQ %08x\n", cqn);
+		mlx4_dbg(dev, "Async event for bogus CQ %08x\n", cqn);
 		return;
 	}
 
+	/* Acessing the CQ outside of rcu_read_lock is safe, because
+	 * the CQ is freed only after interrupt handling is completed.
+	 */
 	cq->event(cq, event_type);
-
-	if (atomic_dec_and_test(&cq->refcount))
-		complete(&cq->free);
 }
 
 static int mlx4_SW2HW_CQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
@@ -216,9 +218,9 @@ int mlx4_cq_alloc(struct mlx4_dev *dev,
 	if (err)
 		goto err_put;
 
-	spin_lock_irq(&cq_table->lock);
+	spin_lock(&cq_table->lock);
 	err = radix_tree_insert(&cq_table->tree, cq->cqn, cq);
-	spin_unlock_irq(&cq_table->lock);
+	spin_unlock(&cq_table->lock);
 	if (err)
 		goto err_cmpt_put;
 
@@ -255,9 +257,9 @@ int mlx4_cq_alloc(struct mlx4_dev *dev,
 	return 0;
 
 err_radix:
-	spin_lock_irq(&cq_table->lock);
+	spin_lock(&cq_table->lock);
 	radix_tree_delete(&cq_table->tree, cq->cqn);
-	spin_unlock_irq(&cq_table->lock);
+	spin_unlock(&cq_table->lock);
 
 err_cmpt_put:
 	mlx4_table_put(dev, &cq_table->cmpt_table, cq->cqn);
@@ -282,11 +284,11 @@ void mlx4_cq_free(struct mlx4_dev *dev,
 	if (err)
 		mlx4_warn(dev, "HW2SW_CQ failed (%d) for CQN %06x\n", err, cq->cqn);
 
-	synchronize_irq(priv->eq_table.eq[cq->vector].irq);
-
-	spin_lock_irq(&cq_table->lock);
+	spin_lock(&cq_table->lock);
 	radix_tree_delete(&cq_table->tree, cq->cqn);
-	spin_unlock_irq(&cq_table->lock);
+	spin_unlock(&cq_table->lock);
+
+	synchronize_irq(priv->eq_table.eq[cq->vector].irq);
 
 	if (atomic_dec_and_test(&cq->refcount))
 		complete(&cq->free);