RDMA/cxgb3: Release dependent resources only when endpoint memory is freed.

The cxgb3 l2t entry, hwtid, and dst entry were being released before
all the iwch_ep references were released.  This can cause a crash in
t3_l2t_send_slow() and other places where the l2t entry is used.

The fix is to defer releasing these resources until all endpoint
references are gone.

Details:

- move flags field to the iwch_ep_common struct.
- add a flag indicating resources are to be released.
- release resources at endpoint free time instead of close/abort time.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

2 files changed, 17 insertions, 12 deletions

diff --git a/drivers/infiniband/hw/cxgb3/iwch_cm.c b/drivers/infiniband/hw/cxgb3/iwch_cm.c
index 59e1c5f0..fef3f1a 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_cm.c
+++ b/drivers/infiniband/hw/cxgb3/iwch_cm.c
@@ -282,18 +282,22 @@ static void *alloc_ep(int size, gfp_t gfp)
 
 void __free_ep(struct kref *kref)
 {
-	struct iwch_ep_common *epc;
-	epc = container_of(kref, struct iwch_ep_common, kref);
-	PDBG("%s ep %p state %s\n", __func__, epc, states[state_read(epc)]);
-	kfree(epc);
+	struct iwch_ep *ep;
+	ep = container_of(container_of(kref, struct iwch_ep_common, kref),
+			  struct iwch_ep, com);
+	PDBG("%s ep %p state %s\n", __func__, ep, states[state_read(&ep->com)]);
+	if (ep->com.flags & RELEASE_RESOURCES) {
+		cxgb3_remove_tid(ep->com.tdev, (void *)ep, ep->hwtid);
+		dst_release(ep->dst);
+		l2t_release(L2DATA(ep->com.tdev), ep->l2t);
+	}
+	kfree(ep);
 }
 
 static void release_ep_resources(struct iwch_ep *ep)
 {
 	PDBG("%s ep %p tid %d\n", __func__, ep, ep->hwtid);
-	cxgb3_remove_tid(ep->com.tdev, (void *)ep, ep->hwtid);
-	dst_release(ep->dst);
-	l2t_release(L2DATA(ep->com.tdev), ep->l2t);
+	ep->com.flags |= RELEASE_RESOURCES;
 	put_ep(&ep->com);
 }
 
@@ -1152,8 +1156,8 @@ static int abort_rpl(struct t3cdev *tdev, struct sk_buff *skb, void *ctx)
 	 * We get 2 abort replies from the HW.  The first one must
 	 * be ignored except for scribbling that we need one more.
 	 */
-	if (!(ep->flags & ABORT_REQ_IN_PROGRESS)) {
-		ep->flags |= ABORT_REQ_IN_PROGRESS;
+	if (!(ep->com.flags & ABORT_REQ_IN_PROGRESS)) {
+		ep->com.flags |= ABORT_REQ_IN_PROGRESS;
 		return CPL_RET_BUF_DONE;
 	}
 
@@ -1557,8 +1561,8 @@ static int peer_abort(struct t3cdev *tdev, struct sk_buff *skb, void *ctx)
 	 * We get 2 peer aborts from the HW.  The first one must
 	 * be ignored except for scribbling that we need one more.
 	 */
-	if (!(ep->flags & PEER_ABORT_IN_PROGRESS)) {
-		ep->flags |= PEER_ABORT_IN_PROGRESS;
+	if (!(ep->com.flags & PEER_ABORT_IN_PROGRESS)) {
+		ep->com.flags |= PEER_ABORT_IN_PROGRESS;
 		return CPL_RET_BUF_DONE;
 	}
 
diff --git a/drivers/infiniband/hw/cxgb3/iwch_cm.h b/drivers/infiniband/hw/cxgb3/iwch_cm.h
index d7c7e09..43c0aea 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_cm.h
+++ b/drivers/infiniband/hw/cxgb3/iwch_cm.h
@@ -147,6 +147,7 @@ enum iwch_ep_state {
 enum iwch_ep_flags {
 	PEER_ABORT_IN_PROGRESS	= (1 << 0),
 	ABORT_REQ_IN_PROGRESS	= (1 << 1),
+	RELEASE_RESOURCES	= (1 << 2),
 };
 
 struct iwch_ep_common {
@@ -161,6 +162,7 @@ struct iwch_ep_common {
 	wait_queue_head_t waitq;
 	int rpl_done;
 	int rpl_err;
+	u32 flags;
 };
 
 struct iwch_listen_ep {
@@ -188,7 +190,6 @@ struct iwch_ep {
 	u16 plen;
 	u32 ird;
 	u32 ord;
-	u32 flags;
 };
 
 static inline struct iwch_ep *to_ep(struct iw_cm_id *cm_id)