[openib-general] Re: [PATCH] mthca: missing cleanup in reset state

[Roland Dreier]

    Michael> The following patch fixes system hangs I am sometimes
    Michael> seeing when ipoib is brought down and back up.

Thanks, applied and queued for 2.6.15.

    Michael> Should something similiar be done in libmthca?

I think so -- like this?

--- libmthca/src/qp.c	(revision 4128)
+++ libmthca/src/qp.c	(working copy)
@@ -65,6 +65,21 @@ static void *get_send_wqe(struct mthca_q
 	return qp->buf + qp->send_wqe_offset + (n << qp->sq.wqe_shift);
 }
 
+void mthca_init_qp_indices(struct mthca_qp *qp)
+{
+	qp->sq.next_ind  = 0;
+	qp->sq.last_comp = qp->sq.max - 1;
+	qp->sq.head    	 = 0;
+	qp->sq.tail    	 = 0;
+	qp->sq.last      = get_send_wqe(qp, qp->sq.max - 1);
+
+	qp->rq.next_ind	 = 0;
+	qp->rq.last_comp = qp->rq.max - 1;
+	qp->rq.head    	 = 0;
+	qp->rq.tail    	 = 0;
+	qp->rq.last      = get_recv_wqe(qp, qp->rq.max - 1);
+}
+
 static inline int wq_overflow(struct mthca_wq *wq, int nreq, struct mthca_cq *cq)
 {
 	unsigned cur;
--- libmthca/src/verbs.c	(revision 4128)
+++ libmthca/src/verbs.c	(working copy)
@@ -394,19 +394,6 @@ int mthca_destroy_srq(struct ibv_srq *sr
 	return 0;
 }
 
-static void mthca_init_qp_indices(struct mthca_qp *qp)
-{
-	qp->sq.next_ind  = 0;
-	qp->sq.last_comp = qp->sq.max - 1;
-	qp->sq.head    	 = 0;
-	qp->sq.tail    	 = 0;
-
-	qp->rq.next_ind	 = 0;
-	qp->rq.last_comp = qp->rq.max - 1;
-	qp->rq.head    	 = 0;
-	qp->rq.tail    	 = 0;
-}
-
 struct ibv_qp *mthca_create_qp(struct ibv_pd *pd, struct ibv_qp_init_attr *attr)
 {
 	struct mthca_create_qp cmd;
@@ -427,11 +414,12 @@ struct ibv_qp *mthca_create_qp(struct ib
 
 	qp->sq.max = align_queue_size(pd->context, attr->cap.max_send_wr, 0);
 	qp->rq.max = align_queue_size(pd->context, attr->cap.max_recv_wr, 0);
-	mthca_init_qp_indices(qp);
 
 	if (mthca_alloc_qp_buf(pd, &attr->cap, attr->qp_type, qp))
 		goto err;
 
+	mthca_init_qp_indices(qp);
+
 	if (pthread_spin_init(&qp->sq.lock, PTHREAD_PROCESS_PRIVATE) ||
 	    pthread_spin_init(&qp->rq.lock, PTHREAD_PROCESS_PRIVATE))
 		goto err_free;
--- libmthca/src/mthca.h	(revision 4128)
+++ libmthca/src/mthca.h	(working copy)
@@ -310,6 +310,7 @@ extern struct ibv_qp *mthca_create_qp(st
 extern int mthca_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
 			   enum ibv_qp_attr_mask attr_mask);
 extern int mthca_destroy_qp(struct ibv_qp *qp);
+extern void mthca_init_qp_indices(struct mthca_qp *qp);
 extern int mthca_tavor_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr,
 				 struct ibv_send_wr **bad_wr);
 extern int mthca_tavor_post_recv(struct ibv_qp *ibqp, struct ibv_recv_wr *wr,
--- libmthca/ChangeLog	(revision 4128)
+++ libmthca/ChangeLog	(working copy)
@@ -1,3 +1,9 @@
+2005-11-28  Roland Dreier  <roland at cisco.com>
+
+	* src/qp.c (mthca_init_qp_indices): Set qp->sq.last and
+	qp->rq.last so that QP is fully reset when the indices are
+	reinited on transition to RESET state.
+
 2005-11-09  Roland Dreier  <roland at cisco.com>
 
 	* src/srq.c (mthca_tavor_post_srq_recv), src/qp.c