diff --git a/drivers/infiniband/hw/cxgb4/cm.c b/drivers/infiniband/hw/cxgb4/cm.c
index 6517113..49784a4 100644
--- a/drivers/infiniband/hw/cxgb4/cm.c
+++ b/drivers/infiniband/hw/cxgb4/cm.c
@@ -145,6 +145,7 @@
 static struct sk_buff *get_skb(struct sk_buff *skb, int len, gfp_t gfp);
 static void ep_timeout(unsigned long arg);
 static void connect_reply_upcall(struct c4iw_ep *ep, int status);
+static int sched(struct c4iw_dev *dev, struct sk_buff *skb);
 
 static LIST_HEAD(timeout_list);
 static spinlock_t timeout_lock;
@@ -295,7 +296,7 @@
 	struct c4iw_ep *ep;
 
 	ep = container_of(kref, struct c4iw_ep, com.kref);
-	PDBG("%s ep %p state %s\n", __func__, ep, states[state_read(&ep->com)]);
+	PDBG("%s ep %p state %s\n", __func__, ep, states[ep->com.state]);
 	if (test_bit(QP_REFERENCED, &ep->com.flags))
 		deref_qp(ep);
 	if (test_bit(RELEASE_RESOURCES, &ep->com.flags)) {
@@ -432,10 +433,57 @@
 
 static void arp_failure_discard(void *handle, struct sk_buff *skb)
 {
-	PDBG("%s c4iw_dev %p\n", __func__, handle);
+	pr_err(MOD "ARP failure\n");
 	kfree_skb(skb);
 }
 
+enum {
+	NUM_FAKE_CPLS = 1,
+	FAKE_CPL_PUT_EP_SAFE = NUM_CPL_CMDS + 0,
+};
+
+static int _put_ep_safe(struct c4iw_dev *dev, struct sk_buff *skb)
+{
+	struct c4iw_ep *ep;
+
+	ep = *((struct c4iw_ep **)(skb->cb + 2 * sizeof(void *)));
+	release_ep_resources(ep);
+	return 0;
+}
+
+/*
+ * Fake up a special CPL opcode and call sched() so process_work() will call
+ * _put_ep_safe() in a safe context to free the ep resources.  This is needed
+ * because ARP error handlers are called in an ATOMIC context, and
+ * _c4iw_free_ep() needs to block.
+ */
+static void queue_arp_failure_cpl(struct c4iw_ep *ep, struct sk_buff *skb)
+{
+	struct cpl_act_establish *rpl = cplhdr(skb);
+
+	/* Set our special ARP_FAILURE opcode */
+	rpl->ot.opcode = FAKE_CPL_PUT_EP_SAFE;
+
+	/*
+	 * Save ep in the skb->cb area, after where sched() will save the dev
+	 * ptr.
+	 */
+	*((struct c4iw_ep **)(skb->cb + 2 * sizeof(void *))) = ep;
+	sched(ep->com.dev, skb);
+}
+
+/* Handle an ARP failure for an accept */
+static void pass_accept_rpl_arp_failure(void *handle, struct sk_buff *skb)
+{
+	struct c4iw_ep *ep = handle;
+
+	pr_err(MOD "ARP failure during accept - tid %u -dropping connection\n",
+	       ep->hwtid);
+
+	__state_set(&ep->com, DEAD);
+	queue_arp_failure_cpl(ep, skb);
+}
+
 /*
  * Handle an ARP failure for an active open.
  */
@@ -444,9 +492,8 @@
 	struct c4iw_ep *ep = handle;
 
 	printk(KERN_ERR MOD "ARP failure during connect\n");
-	kfree_skb(skb);
 	connect_reply_upcall(ep, -EHOSTUNREACH);
-	state_set(&ep->com, DEAD);
+	__state_set(&ep->com, DEAD);
 	if (ep->com.remote_addr.ss_family == AF_INET6) {
 		struct sockaddr_in6 *sin6 =
 			(struct sockaddr_in6 *)&ep->com.local_addr;
@@ -455,9 +502,7 @@
 	}
 	remove_handle(ep->com.dev, &ep->com.dev->atid_idr, ep->atid);
 	cxgb4_free_atid(ep->com.dev->rdev.lldi.tids, ep->atid);
-	dst_release(ep->dst);
-	cxgb4_l2t_release(ep->l2t);
-	c4iw_put_ep(&ep->com);
+	queue_arp_failure_cpl(ep, skb);
 }
 
 /*
@@ -2198,8 +2243,8 @@
 	return 0;
 }
 
-static void accept_cr(struct c4iw_ep *ep, struct sk_buff *skb,
-		      struct cpl_pass_accept_req *req)
+static int accept_cr(struct c4iw_ep *ep, struct sk_buff *skb,
+		     struct cpl_pass_accept_req *req)
 {
 	struct cpl_pass_accept_rpl *rpl;
 	unsigned int mtu_idx;
@@ -2287,10 +2332,9 @@
 	rpl->opt0 = cpu_to_be64(opt0);
 	rpl->opt2 = cpu_to_be32(opt2);
 	set_wr_txq(skb, CPL_PRIORITY_SETUP, ep->ctrlq_idx);
-	t4_set_arp_err_handler(skb, NULL, arp_failure_discard);
-	c4iw_l2t_send(&ep->com.dev->rdev, skb, ep->l2t);
+	t4_set_arp_err_handler(skb, ep, pass_accept_rpl_arp_failure);
 
-	return;
+	return c4iw_l2t_send(&ep->com.dev->rdev, skb, ep->l2t);
 }
 
 static void reject_cr(struct c4iw_dev *dev, u32 hwtid, struct sk_buff *skb)
@@ -2469,8 +2513,12 @@
 	init_timer(&child_ep->timer);
 	cxgb4_insert_tid(t, child_ep, hwtid);
 	insert_handle(dev, &dev->hwtid_idr, child_ep, child_ep->hwtid);
-	accept_cr(child_ep, skb, req);
-	set_bit(PASS_ACCEPT_REQ, &child_ep->com.history);
+	if (accept_cr(child_ep, skb, req)) {
+		c4iw_put_ep(&parent_ep->com);
+		release_ep_resources(child_ep);
+	} else {
+		set_bit(PASS_ACCEPT_REQ, &child_ep->com.history);
+	}
 	if (iptype == 6) {
 		sin6 = (struct sockaddr_in6 *)&child_ep->com.local_addr;
 		cxgb4_clip_get(child_ep->com.dev->rdev.lldi.ports[0],
@@ -2633,6 +2681,7 @@
 	mutex_lock(&ep->com.mutex);
 	switch (ep->com.state) {
 	case CONNECTING:
+		c4iw_put_ep(&ep->parent_ep->com);
 		break;
 	case MPA_REQ_WAIT:
 		(void)stop_ep_timer(ep);
@@ -3809,7 +3858,7 @@
  * These are the real handlers that are called from a
  * work queue.
  */
-static c4iw_handler_func work_handlers[NUM_CPL_CMDS] = {
+static c4iw_handler_func work_handlers[NUM_CPL_CMDS + NUM_FAKE_CPLS] = {
 	[CPL_ACT_ESTABLISH] = act_establish,
 	[CPL_ACT_OPEN_RPL] = act_open_rpl,
 	[CPL_RX_DATA] = rx_data,
@@ -3825,7 +3874,8 @@
 	[CPL_RDMA_TERMINATE] = terminate,
 	[CPL_FW4_ACK] = fw4_ack,
 	[CPL_FW6_MSG] = deferred_fw6_msg,
-	[CPL_RX_PKT] = rx_pkt
+	[CPL_RX_PKT] = rx_pkt,
+	[FAKE_CPL_PUT_EP_SAFE] = _put_ep_safe
 };
 
 static void process_timeout(struct c4iw_ep *ep)