IB/mthca: Fix race in reference counting

Fix races in in destroying various objects.  If a destroy routine
waits for an object to become free by doing

	wait_event(&obj->wait, !atomic_read(&obj->refcount));
	/* now clean up and destroy the object */

and another place drops a reference to the object by doing

	if (atomic_dec_and_test(&obj->refcount))
		wake_up(&obj->wait);

then this is susceptible to a race where the wait_event() and final
freeing of the object occur between the atomic_dec_and_test() and the
wake_up().  And this is a use-after-free, since wake_up() will be
called on part of the already-freed object.

Fix this in mthca by replacing the atomic_t refcounts with plain old
integers protected by a spinlock.  This makes it possible to do the
decrement of the reference count and the wake_up() so that it appears
as a single atomic operation to the code waiting on the wait queue.

While touching this code, also simplify mthca_cq_clean(): the CQ being
cleaned cannot go away, because it still has a QP attached to it.  So
there's no reason to be paranoid and look up the CQ by number; it's
perfectly safe to use the pointer that the callers already have.

Signed-off-by: Roland Dreier <rolandd@cisco.com>

5 files changed, 74 insertions, 45 deletions

diff --git a/drivers/infiniband/hw/mthca/mthca_cq.c b/drivers/infiniband/hw/mthca/mthca_cq.c
index 312cf90731ea..205854e9c662 100644
--- a/drivers/infiniband/hw/mthca/mthca_cq.c
+++ b/drivers/infiniband/hw/mthca/mthca_cq.c
@@ -238,9 +238,9 @@ void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
 	spin_lock(&dev->cq_table.lock);
 
 	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
-
 	if (cq)
-		atomic_inc(&cq->refcount);
+		++cq->refcount;
+
 	spin_unlock(&dev->cq_table.lock);
 
 	if (!cq) {
@@ -254,8 +254,10 @@ void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
 	if (cq->ibcq.event_handler)
 		cq->ibcq.event_handler(&event, cq->ibcq.cq_context);
 
-	if (atomic_dec_and_test(&cq->refcount))
+	spin_lock(&dev->cq_table.lock);
+	if (!--cq->refcount)
 		wake_up(&cq->wait);
+	spin_unlock(&dev->cq_table.lock);
 }
 
 static inline int is_recv_cqe(struct mthca_cqe *cqe)
@@ -267,23 +269,13 @@ static inline int is_recv_cqe(struct mthca_cqe *cqe)
 		return !(cqe->is_send & 0x80);
 }
 
-void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
+void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
 		    struct mthca_srq *srq)
 {
-	struct mthca_cq *cq;
 	struct mthca_cqe *cqe;
 	u32 prod_index;
 	int nfreed = 0;
 
-	spin_lock_irq(&dev->cq_table.lock);
-	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
-	if (cq)
-		atomic_inc(&cq->refcount);
-	spin_unlock_irq(&dev->cq_table.lock);
-
-	if (!cq)
-		return;
-
 	spin_lock_irq(&cq->lock);
 
 	/*
@@ -301,7 +293,7 @@ void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
 
 	if (0)
 		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
-			  qpn, cqn, cq->cons_index, prod_index);
+			  qpn, cq->cqn, cq->cons_index, prod_index);
 
 	/*
 	 * Now sweep backwards through the CQ, removing CQ entries
@@ -325,8 +317,6 @@ void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
 	}
 
 	spin_unlock_irq(&cq->lock);
-	if (atomic_dec_and_test(&cq->refcount))
-		wake_up(&cq->wait);
 }
 
 void mthca_cq_resize_copy_cqes(struct mthca_cq *cq)
@@ -821,7 +811,7 @@ int mthca_init_cq(struct mthca_dev *dev, int nent,
 	}
 
 	spin_lock_init(&cq->lock);
-	atomic_set(&cq->refcount, 1);
+	cq->refcount = 1;
 	init_waitqueue_head(&cq->wait);
 
 	memset(cq_context, 0, sizeof *cq_context);
@@ -896,6 +886,17 @@ err_out:
 	return err;
 }
 
+static inline int get_cq_refcount(struct mthca_dev *dev, struct mthca_cq *cq)
+{
+	int c;
+
+	spin_lock_irq(&dev->cq_table.lock);
+	c = cq->refcount;
+	spin_unlock_irq(&dev->cq_table.lock);
+
+	return c;
+}
+
 void mthca_free_cq(struct mthca_dev *dev,
 		   struct mthca_cq *cq)
 {
@@ -929,6 +930,7 @@ void mthca_free_cq(struct mthca_dev *dev,
 	spin_lock_irq(&dev->cq_table.lock);
 	mthca_array_clear(&dev->cq_table.cq,
 			  cq->cqn & (dev->limits.num_cqs - 1));
+	--cq->refcount;
 	spin_unlock_irq(&dev->cq_table.lock);
 
 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
@@ -936,8 +938,7 @@ void mthca_free_cq(struct mthca_dev *dev,
 	else
 		synchronize_irq(dev->pdev->irq);
 
-	atomic_dec(&cq->refcount);
-	wait_event(cq->wait, !atomic_read(&cq->refcount));
+	wait_event(cq->wait, !get_cq_refcount(dev, cq));
 
 	if (cq->is_kernel) {
 		mthca_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
diff --git a/drivers/infiniband/hw/mthca/mthca_dev.h b/drivers/infiniband/hw/mthca/mthca_dev.h
index 4c1dcb4c1822..f8160b8de090 100644
--- a/drivers/infiniband/hw/mthca/mthca_dev.h
+++ b/drivers/infiniband/hw/mthca/mthca_dev.h
@@ -496,7 +496,7 @@ void mthca_free_cq(struct mthca_dev *dev,
 void mthca_cq_completion(struct mthca_dev *dev, u32 cqn);
 void mthca_cq_event(struct mthca_dev *dev, u32 cqn,
 		    enum ib_event_type event_type);
-void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn,
+void mthca_cq_clean(struct mthca_dev *dev, struct mthca_cq *cq, u32 qpn,
 		    struct mthca_srq *srq);
 void mthca_cq_resize_copy_cqes(struct mthca_cq *cq);
 int mthca_alloc_cq_buf(struct mthca_dev *dev, struct mthca_cq_buf *buf, int nent);
diff --git a/drivers/infiniband/hw/mthca/mthca_provider.h b/drivers/infiniband/hw/mthca/mthca_provider.h
index 6676a786d690..179a8f610d0f 100644
--- a/drivers/infiniband/hw/mthca/mthca_provider.h
+++ b/drivers/infiniband/hw/mthca/mthca_provider.h
@@ -139,11 +139,12 @@ struct mthca_ah {
  * a qp may be locked, with the send cq locked first.  No other
  * nesting should be done.
  *
- * Each struct mthca_cq/qp also has an atomic_t ref count.  The
- * pointer from the cq/qp_table to the struct counts as one reference.
- * This reference also is good for access through the consumer API, so
- * modifying the CQ/QP etc doesn't need to take another reference.
- * Access because of a completion being polled does need a reference.
+ * Each struct mthca_cq/qp also has an ref count, protected by the
+ * corresponding table lock.  The pointer from the cq/qp_table to the
+ * struct counts as one reference.  This reference also is good for
+ * access through the consumer API, so modifying the CQ/QP etc doesn't
+ * need to take another reference.  Access to a QP because of a
+ * completion being polled does not need a reference either.
  *
  * Finally, each struct mthca_cq/qp has a wait_queue_head_t for the
  * destroy function to sleep on.
@@ -159,8 +160,9 @@ struct mthca_ah {
  * - decrement ref count; if zero, wake up waiters
  *
  * To destroy a CQ/QP, we can do the following:
- * - lock cq/qp_table, remove pointer, unlock cq/qp_table lock
- * - decrement ref count
+ * - lock cq/qp_table
+ * - remove pointer and decrement ref count
+ * - unlock cq/qp_table lock
  * - wait_event until ref count is zero
  *
  * It is the consumer's responsibilty to make sure that no QP
@@ -197,7 +199,7 @@ struct mthca_cq_resize {
 struct mthca_cq {
 	struct ib_cq		ibcq;
 	spinlock_t		lock;
-	atomic_t		refcount;
+	int			refcount;
 	int			cqn;
 	u32			cons_index;
 	struct mthca_cq_buf	buf;
@@ -217,7 +219,7 @@ struct mthca_cq {
 struct mthca_srq {
 	struct ib_srq		ibsrq;
 	spinlock_t		lock;
-	atomic_t		refcount;
+	int			refcount;
 	int			srqn;
 	int			max;
 	int			max_gs;
@@ -254,7 +256,7 @@ struct mthca_wq {
 
 struct mthca_qp {
 	struct ib_qp           ibqp;
-	atomic_t               refcount;
+	int                    refcount;
 	u32                    qpn;
 	int                    is_direct;
 	u8                     port; /* for SQP and memfree use only */
diff --git a/drivers/infiniband/hw/mthca/mthca_qp.c b/drivers/infiniband/hw/mthca/mthca_qp.c
index f37b0e367323..19765f6f8d58 100644
--- a/drivers/infiniband/hw/mthca/mthca_qp.c
+++ b/drivers/infiniband/hw/mthca/mthca_qp.c
@@ -240,7 +240,7 @@ void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
 	spin_lock(&dev->qp_table.lock);
 	qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
 	if (qp)
-		atomic_inc(&qp->refcount);
+		++qp->refcount;
 	spin_unlock(&dev->qp_table.lock);
 
 	if (!qp) {
@@ -257,8 +257,10 @@ void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
 	if (qp->ibqp.event_handler)
 		qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
 
-	if (atomic_dec_and_test(&qp->refcount))
+	spin_lock(&dev->qp_table.lock);
+	if (!--qp->refcount)
 		wake_up(&qp->wait);
+	spin_unlock(&dev->qp_table.lock);
 }
 
 static int to_mthca_state(enum ib_qp_state ib_state)
@@ -833,10 +835,10 @@ int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
 	 * entries and reinitialize the QP.
 	 */
 	if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
-		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
+		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn,
 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
-			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
+			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
 				       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 
 		mthca_wq_init(&qp->sq);
@@ -1096,7 +1098,7 @@ static int mthca_alloc_qp_common(struct mthca_dev *dev,
 	int ret;
 	int i;
 
-	atomic_set(&qp->refcount, 1);
+	qp->refcount = 1;
 	init_waitqueue_head(&qp->wait);
 	qp->state    	 = IB_QPS_RESET;
 	qp->atomic_rd_en = 0;
@@ -1318,6 +1320,17 @@ int mthca_alloc_sqp(struct mthca_dev *dev,
 	return err;
 }
 
+static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
+{
+	int c;
+
+	spin_lock_irq(&dev->qp_table.lock);
+	c = qp->refcount;
+	spin_unlock_irq(&dev->qp_table.lock);
+
+	return c;
+}
+
 void mthca_free_qp(struct mthca_dev *dev,
 		   struct mthca_qp *qp)
 {
@@ -1339,14 +1352,14 @@ void mthca_free_qp(struct mthca_dev *dev,
 	spin_lock(&dev->qp_table.lock);
 	mthca_array_clear(&dev->qp_table.qp,
 			  qp->qpn & (dev->limits.num_qps - 1));
+	--qp->refcount;
 	spin_unlock(&dev->qp_table.lock);
 
 	if (send_cq != recv_cq)
 		spin_unlock(&recv_cq->lock);
 	spin_unlock_irq(&send_cq->lock);
 
-	atomic_dec(&qp->refcount);
-	wait_event(qp->wait, !atomic_read(&qp->refcount));
+	wait_event(qp->wait, !get_qp_refcount(dev, qp));
 
 	if (qp->state != IB_QPS_RESET)
 		mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
@@ -1358,10 +1371,10 @@ void mthca_free_qp(struct mthca_dev *dev,
 	 * unref the mem-free tables and free the QPN in our table.
 	 */
 	if (!qp->ibqp.uobject) {
-		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
+		mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn,
 			       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 		if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
-			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
+			mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
 				       qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
 
 		mthca_free_memfree(dev, qp);
diff --git a/drivers/infiniband/hw/mthca/mthca_srq.c b/drivers/infiniband/hw/mthca/mthca_srq.c
index adcaf85355ae..1ea433291fa7 100644
--- a/drivers/infiniband/hw/mthca/mthca_srq.c
+++ b/drivers/infiniband/hw/mthca/mthca_srq.c
@@ -241,7 +241,7 @@ int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
 		goto err_out_mailbox;
 
 	spin_lock_init(&srq->lock);
-	atomic_set(&srq->refcount, 1);
+	srq->refcount = 1;
 	init_waitqueue_head(&srq->wait);
 
 	if (mthca_is_memfree(dev))
@@ -308,6 +308,17 @@ err_out:
 	return err;
 }
 
+static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
+{
+	int c;
+
+	spin_lock_irq(&dev->srq_table.lock);
+	c = srq->refcount;
+	spin_unlock_irq(&dev->srq_table.lock);
+
+	return c;
+}
+
 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
 {
 	struct mthca_mailbox *mailbox;
@@ -329,10 +340,10 @@ void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
 	spin_lock_irq(&dev->srq_table.lock);
 	mthca_array_clear(&dev->srq_table.srq,
 			  srq->srqn & (dev->limits.num_srqs - 1));
+	--srq->refcount;
 	spin_unlock_irq(&dev->srq_table.lock);
 
-	atomic_dec(&srq->refcount);
-	wait_event(srq->wait, !atomic_read(&srq->refcount));
+	wait_event(srq->wait, !get_srq_refcount(dev, srq));
 
 	if (!srq->ibsrq.uobject) {
 		mthca_free_srq_buf(dev, srq);
@@ -414,7 +425,7 @@ void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
 	spin_lock(&dev->srq_table.lock);
 	srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
 	if (srq)
-		atomic_inc(&srq->refcount);
+		++srq->refcount;
 	spin_unlock(&dev->srq_table.lock);
 
 	if (!srq) {
@@ -431,8 +442,10 @@ void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
 	srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
 
 out:
-	if (atomic_dec_and_test(&srq->refcount))
+	spin_lock(&dev->srq_table.lock);
+	if (!--srq->refcount)
 		wake_up(&srq->wait);
+	spin_unlock(&dev->srq_table.lock);
 }
 
 /*