Patchwork ib/ehca: add flush CQE generation

login
register
mail settings
Submitter Alexander Schmidt
Date Sept. 10, 2008, 2:23 p.m.
Message ID <20080910162356.7294fe87@BL3D1974.boeblingen.de.ibm.com>
Download mbox | patch
Permalink /patch/227/
State Not Applicable, archived
Headers show

Comments

Alexander Schmidt - Sept. 10, 2008, 2:23 p.m.
When a QP goes into error state, it is required that flush CQEs are
delivered to the application for any outstanding work requests. eHCA does not
do this in hardware, so this patch adds software flush CQE generation to the
ehca driver.

Whenever a QP gets into error state, it is added to the QP error list of its
respective CQ. If the error QP list of a CQ is not empty, poll_cq()
generates flush CQEs before polling the actual CQ.

Signed-off-by: Alexander Schmidt <alexs@linux.vnet.ibm.com>
Roland Dreier - Sept. 16, 2008, 4:45 p.m.
thanks, queued for 2.6.28.

Patch

--- infiniband.git.orig/drivers/infiniband/hw/ehca/ehca_classes.h
+++ infiniband.git/drivers/infiniband/hw/ehca/ehca_classes.h
@@ -164,6 +164,13 @@  struct ehca_qmap_entry {
 	u16 reported;
 };
 
+struct ehca_queue_map {
+	struct ehca_qmap_entry *map;
+	unsigned int entries;
+	unsigned int tail;
+	unsigned int left_to_poll;
+};
+
 struct ehca_qp {
 	union {
 		struct ib_qp ib_qp;
@@ -173,8 +180,9 @@  struct ehca_qp {
 	enum ehca_ext_qp_type ext_type;
 	enum ib_qp_state state;
 	struct ipz_queue ipz_squeue;
-	struct ehca_qmap_entry *sq_map;
+	struct ehca_queue_map sq_map;
 	struct ipz_queue ipz_rqueue;
+	struct ehca_queue_map rq_map;
 	struct h_galpas galpas;
 	u32 qkey;
 	u32 real_qp_num;
@@ -204,6 +212,8 @@  struct ehca_qp {
 	atomic_t nr_events; /* events seen */
 	wait_queue_head_t wait_completion;
 	int mig_armed;
+	struct list_head sq_err_node;
+	struct list_head rq_err_node;
 };
 
 #define IS_SRQ(qp) (qp->ext_type == EQPT_SRQ)
@@ -233,6 +243,8 @@  struct ehca_cq {
 	/* mmap counter for resources mapped into user space */
 	u32 mm_count_queue;
 	u32 mm_count_galpa;
+	struct list_head sqp_err_list;
+	struct list_head rqp_err_list;
 };
 
 enum ehca_mr_flag {
--- infiniband.git.orig/drivers/infiniband/hw/ehca/ehca_reqs.c
+++ infiniband.git/drivers/infiniband/hw/ehca/ehca_reqs.c
@@ -53,9 +53,25 @@ 
 /* in RC traffic, insert an empty RDMA READ every this many packets */
 #define ACK_CIRC_THRESHOLD 2000000
 
+static u64 replace_wr_id(u64 wr_id, u16 idx)
+{
+	u64 ret;
+
+	ret = wr_id & ~QMAP_IDX_MASK;
+	ret |= idx & QMAP_IDX_MASK;
+
+	return ret;
+}
+
+static u16 get_app_wr_id(u64 wr_id)
+{
+	return wr_id & QMAP_IDX_MASK;
+}
+
 static inline int ehca_write_rwqe(struct ipz_queue *ipz_rqueue,
 				  struct ehca_wqe *wqe_p,
-				  struct ib_recv_wr *recv_wr)
+				  struct ib_recv_wr *recv_wr,
+				  u32 rq_map_idx)
 {
 	u8 cnt_ds;
 	if (unlikely((recv_wr->num_sge < 0) ||
@@ -69,7 +85,7 @@  static inline int ehca_write_rwqe(struct
 	/* clear wqe header until sglist */
 	memset(wqe_p, 0, offsetof(struct ehca_wqe, u.ud_av.sg_list));
 
-	wqe_p->work_request_id = recv_wr->wr_id;
+	wqe_p->work_request_id = replace_wr_id(recv_wr->wr_id, rq_map_idx);
 	wqe_p->nr_of_data_seg = recv_wr->num_sge;
 
 	for (cnt_ds = 0; cnt_ds < recv_wr->num_sge; cnt_ds++) {
@@ -146,6 +162,7 @@  static inline int ehca_write_swqe(struct
 	u64 dma_length;
 	struct ehca_av *my_av;
 	u32 remote_qkey = send_wr->wr.ud.remote_qkey;
+	struct ehca_qmap_entry *qmap_entry = &qp->sq_map.map[sq_map_idx];
 
 	if (unlikely((send_wr->num_sge < 0) ||
 		     (send_wr->num_sge > qp->ipz_squeue.act_nr_of_sg))) {
@@ -158,11 +175,10 @@  static inline int ehca_write_swqe(struct
 	/* clear wqe header until sglist */
 	memset(wqe_p, 0, offsetof(struct ehca_wqe, u.ud_av.sg_list));
 
-	wqe_p->work_request_id = send_wr->wr_id & ~QMAP_IDX_MASK;
-	wqe_p->work_request_id |= sq_map_idx & QMAP_IDX_MASK;
+	wqe_p->work_request_id = replace_wr_id(send_wr->wr_id, sq_map_idx);
 
-	qp->sq_map[sq_map_idx].app_wr_id = send_wr->wr_id & QMAP_IDX_MASK;
-	qp->sq_map[sq_map_idx].reported = 0;
+	qmap_entry->app_wr_id = get_app_wr_id(send_wr->wr_id);
+	qmap_entry->reported = 0;
 
 	switch (send_wr->opcode) {
 	case IB_WR_SEND:
@@ -496,7 +512,9 @@  static int internal_post_recv(struct ehc
 	struct ehca_wqe *wqe_p;
 	int wqe_cnt = 0;
 	int ret = 0;
+	u32 rq_map_idx;
 	unsigned long flags;
+	struct ehca_qmap_entry *qmap_entry;
 
 	if (unlikely(!HAS_RQ(my_qp))) {
 		ehca_err(dev, "QP has no RQ  ehca_qp=%p qp_num=%x ext_type=%d",
@@ -524,8 +542,15 @@  static int internal_post_recv(struct ehc
 			}
 			goto post_recv_exit0;
 		}
+		/*
+		 * Get the index of the WQE in the recv queue. The same index
+		 * is used for writing into the rq_map.
+		 */
+		rq_map_idx = start_offset / my_qp->ipz_rqueue.qe_size;
+
 		/* write a RECV WQE into the QUEUE */
-		ret = ehca_write_rwqe(&my_qp->ipz_rqueue, wqe_p, cur_recv_wr);
+		ret = ehca_write_rwqe(&my_qp->ipz_rqueue, wqe_p, cur_recv_wr,
+				rq_map_idx);
 		/*
 		 * if something failed,
 		 * reset the free entry pointer to the start value
@@ -540,6 +565,11 @@  static int internal_post_recv(struct ehc
 			}
 			goto post_recv_exit0;
 		}
+
+		qmap_entry = &my_qp->rq_map.map[rq_map_idx];
+		qmap_entry->app_wr_id = get_app_wr_id(cur_recv_wr->wr_id);
+		qmap_entry->reported = 0;
+
 		wqe_cnt++;
 	} /* eof for cur_recv_wr */
 
@@ -596,10 +626,12 @@  static const u8 ib_wc_opcode[255] = {
 /* internal function to poll one entry of cq */
 static inline int ehca_poll_cq_one(struct ib_cq *cq, struct ib_wc *wc)
 {
-	int ret = 0;
+	int ret = 0, qmap_tail_idx;
 	struct ehca_cq *my_cq = container_of(cq, struct ehca_cq, ib_cq);
 	struct ehca_cqe *cqe;
 	struct ehca_qp *my_qp;
+	struct ehca_qmap_entry *qmap_entry;
+	struct ehca_queue_map *qmap;
 	int cqe_count = 0, is_error;
 
 repoll:
@@ -674,27 +706,52 @@  repoll:
 		goto repoll;
 	wc->qp = &my_qp->ib_qp;
 
-	if (!(cqe->w_completion_flags & WC_SEND_RECEIVE_BIT)) {
-		struct ehca_qmap_entry *qmap_entry;
+	if (is_error) {
 		/*
-		 * We got a send completion and need to restore the original
-		 * wr_id.
+		 * set left_to_poll to 0 because in error state, we will not
+		 * get any additional CQEs
 		 */
-		qmap_entry = &my_qp->sq_map[cqe->work_request_id &
-					    QMAP_IDX_MASK];
+		ehca_add_to_err_list(my_qp, 1);
+		my_qp->sq_map.left_to_poll = 0;
 
-		if (qmap_entry->reported) {
-			ehca_warn(cq->device, "Double cqe on qp_num=%#x",
-				  my_qp->real_qp_num);
-			/* found a double cqe, discard it and read next one */
-			goto repoll;
-		}
-		wc->wr_id = cqe->work_request_id & ~QMAP_IDX_MASK;
-		wc->wr_id |= qmap_entry->app_wr_id;
-		qmap_entry->reported = 1;
-	} else
+		if (HAS_RQ(my_qp))
+			ehca_add_to_err_list(my_qp, 0);
+		my_qp->rq_map.left_to_poll = 0;
+	}
+
+	qmap_tail_idx = get_app_wr_id(cqe->work_request_id);
+	if (!(cqe->w_completion_flags & WC_SEND_RECEIVE_BIT))
+		/* We got a send completion. */
+		qmap = &my_qp->sq_map;
+	else
 		/* We got a receive completion. */
-		wc->wr_id = cqe->work_request_id;
+		qmap = &my_qp->rq_map;
+
+	qmap_entry = &qmap->map[qmap_tail_idx];
+	if (qmap_entry->reported) {
+		ehca_warn(cq->device, "Double cqe on qp_num=%#x",
+				my_qp->real_qp_num);
+		/* found a double cqe, discard it and read next one */
+		goto repoll;
+	}
+
+	wc->wr_id = replace_wr_id(cqe->work_request_id, qmap_entry->app_wr_id);
+	qmap_entry->reported = 1;
+
+	/* this is a proper completion, we need to advance the tail pointer */
+	if (++qmap->tail == qmap->entries)
+		qmap->tail = 0;
+
+	/* if left_to_poll is decremented to 0, add the QP to the error list */
+	if (qmap->left_to_poll > 0) {
+		qmap->left_to_poll--;
+		if ((my_qp->sq_map.left_to_poll == 0) &&
+				(my_qp->rq_map.left_to_poll == 0)) {
+			ehca_add_to_err_list(my_qp, 1);
+			if (HAS_RQ(my_qp))
+				ehca_add_to_err_list(my_qp, 0);
+		}
+	}
 
 	/* eval ib_wc_opcode */
 	wc->opcode = ib_wc_opcode[cqe->optype]-1;
@@ -733,13 +790,88 @@  poll_cq_one_exit0:
 	return ret;
 }
 
+static int generate_flush_cqes(struct ehca_qp *my_qp, struct ib_cq *cq,
+			       struct ib_wc *wc, int num_entries,
+			       struct ipz_queue *ipz_queue, int on_sq)
+{
+	int nr = 0;
+	struct ehca_wqe *wqe;
+	u64 offset;
+	struct ehca_queue_map *qmap;
+	struct ehca_qmap_entry *qmap_entry;
+
+	if (on_sq)
+		qmap = &my_qp->sq_map;
+	else
+		qmap = &my_qp->rq_map;
+
+	qmap_entry = &qmap->map[qmap->tail];
+
+	while ((nr < num_entries) && (qmap_entry->reported == 0)) {
+		/* generate flush CQE */
+		memset(wc, 0, sizeof(*wc));
+
+		offset = qmap->tail * ipz_queue->qe_size;
+		wqe = (struct ehca_wqe *)ipz_qeit_calc(ipz_queue, offset);
+		if (!wqe) {
+			ehca_err(cq->device, "Invalid wqe offset=%#lx on "
+				 "qp_num=%#x", offset, my_qp->real_qp_num);
+			return nr;
+		}
+
+		wc->wr_id = replace_wr_id(wqe->work_request_id,
+					  qmap_entry->app_wr_id);
+
+		if (on_sq) {
+			switch (wqe->optype) {
+			case WQE_OPTYPE_SEND:
+				wc->opcode = IB_WC_SEND;
+				break;
+			case WQE_OPTYPE_RDMAWRITE:
+				wc->opcode = IB_WC_RDMA_WRITE;
+				break;
+			case WQE_OPTYPE_RDMAREAD:
+				wc->opcode = IB_WC_RDMA_READ;
+				break;
+			default:
+				ehca_err(cq->device, "Invalid optype=%x",
+						wqe->optype);
+				return nr;
+			}
+		} else
+			wc->opcode = IB_WC_RECV;
+
+		if (wqe->wr_flag & WQE_WRFLAG_IMM_DATA_PRESENT) {
+			wc->ex.imm_data = wqe->immediate_data;
+			wc->wc_flags |= IB_WC_WITH_IMM;
+		}
+
+		wc->status = IB_WC_WR_FLUSH_ERR;
+
+		wc->qp = &my_qp->ib_qp;
+
+		/* mark as reported and advance tail pointer */
+		qmap_entry->reported = 1;
+		if (++qmap->tail == qmap->entries)
+			qmap->tail = 0;
+		qmap_entry = &qmap->map[qmap->tail];
+
+		wc++; nr++;
+	}
+
+	return nr;
+
+}
+
 int ehca_poll_cq(struct ib_cq *cq, int num_entries, struct ib_wc *wc)
 {
 	struct ehca_cq *my_cq = container_of(cq, struct ehca_cq, ib_cq);
 	int nr;
+	struct ehca_qp *err_qp;
 	struct ib_wc *current_wc = wc;
 	int ret = 0;
 	unsigned long flags;
+	int entries_left = num_entries;
 
 	if (num_entries < 1) {
 		ehca_err(cq->device, "Invalid num_entries=%d ehca_cq=%p "
@@ -749,15 +881,40 @@  int ehca_poll_cq(struct ib_cq *cq, int n
 	}
 
 	spin_lock_irqsave(&my_cq->spinlock, flags);
-	for (nr = 0; nr < num_entries; nr++) {
+
+	/* generate flush cqes for send queues */
+	list_for_each_entry(err_qp, &my_cq->sqp_err_list, sq_err_node) {
+		nr = generate_flush_cqes(err_qp, cq, current_wc, entries_left,
+				&err_qp->ipz_squeue, 1);
+		entries_left -= nr;
+		current_wc += nr;
+
+		if (entries_left == 0)
+			break;
+	}
+
+	/* generate flush cqes for receive queues */
+	list_for_each_entry(err_qp, &my_cq->rqp_err_list, rq_err_node) {
+		nr = generate_flush_cqes(err_qp, cq, current_wc, entries_left,
+				&err_qp->ipz_rqueue, 0);
+		entries_left -= nr;
+		current_wc += nr;
+
+		if (entries_left == 0)
+			break;
+	}
+
+	for (nr = 0; nr < entries_left; nr++) {
 		ret = ehca_poll_cq_one(cq, current_wc);
 		if (ret)
 			break;
 		current_wc++;
 	} /* eof for nr */
+	entries_left -= nr;
+
 	spin_unlock_irqrestore(&my_cq->spinlock, flags);
 	if (ret == -EAGAIN  || !ret)
-		ret = nr;
+		ret = num_entries - entries_left;
 
 poll_cq_exit0:
 	return ret;
--- infiniband.git.orig/drivers/infiniband/hw/ehca/ehca_cq.c
+++ infiniband.git/drivers/infiniband/hw/ehca/ehca_cq.c
@@ -276,6 +276,9 @@  struct ib_cq *ehca_create_cq(struct ib_d
 	for (i = 0; i < QP_HASHTAB_LEN; i++)
 		INIT_HLIST_HEAD(&my_cq->qp_hashtab[i]);
 
+	INIT_LIST_HEAD(&my_cq->sqp_err_list);
+	INIT_LIST_HEAD(&my_cq->rqp_err_list);
+
 	if (context) {
 		struct ipz_queue *ipz_queue = &my_cq->ipz_queue;
 		struct ehca_create_cq_resp resp;
--- infiniband.git.orig/drivers/infiniband/hw/ehca/ehca_qp.c
+++ infiniband.git/drivers/infiniband/hw/ehca/ehca_qp.c
@@ -396,6 +396,50 @@  static void ehca_determine_small_queue(s
 	queue->is_small = (queue->page_size != 0);
 }
 
+/* needs to be called with cq->spinlock held */
+void ehca_add_to_err_list(struct ehca_qp *qp, int on_sq)
+{
+	struct list_head *list, *node;
+
+	/* TODO: support low latency QPs */
+	if (qp->ext_type == EQPT_LLQP)
+		return;
+
+	if (on_sq) {
+		list = &qp->send_cq->sqp_err_list;
+		node = &qp->sq_err_node;
+	} else {
+		list = &qp->recv_cq->rqp_err_list;
+		node = &qp->rq_err_node;
+	}
+
+	if (list_empty(node))
+		list_add_tail(node, list);
+
+	return;
+}
+
+static void del_from_err_list(struct ehca_cq *cq, struct list_head *node)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&cq->spinlock, flags);
+
+	if (!list_empty(node))
+		list_del_init(node);
+
+	spin_unlock_irqrestore(&cq->spinlock, flags);
+}
+
+static void reset_queue_map(struct ehca_queue_map *qmap)
+{
+	int i;
+
+	qmap->tail = 0;
+	for (i = 0; i < qmap->entries; i++)
+		qmap->map[i].reported = 1;
+}
+
 /*
  * Create an ib_qp struct that is either a QP or an SRQ, depending on
  * the value of the is_srq parameter. If init_attr and srq_init_attr share
@@ -407,12 +451,11 @@  static struct ehca_qp *internal_create_q
 	struct ib_srq_init_attr *srq_init_attr,
 	struct ib_udata *udata, int is_srq)
 {
-	struct ehca_qp *my_qp;
+	struct ehca_qp *my_qp, *my_srq = NULL;
 	struct ehca_pd *my_pd = container_of(pd, struct ehca_pd, ib_pd);
 	struct ehca_shca *shca = container_of(pd->device, struct ehca_shca,
 					      ib_device);
 	struct ib_ucontext *context = NULL;
-	u32 nr_qes;
 	u64 h_ret;
 	int is_llqp = 0, has_srq = 0;
 	int qp_type, max_send_sge, max_recv_sge, ret;
@@ -457,8 +500,7 @@  static struct ehca_qp *internal_create_q
 
 	/* handle SRQ base QPs */
 	if (init_attr->srq) {
-		struct ehca_qp *my_srq =
-			container_of(init_attr->srq, struct ehca_qp, ib_srq);
+		my_srq = container_of(init_attr->srq, struct ehca_qp, ib_srq);
 
 		has_srq = 1;
 		parms.ext_type = EQPT_SRQBASE;
@@ -716,15 +758,19 @@  static struct ehca_qp *internal_create_q
 				 "and pages ret=%i", ret);
 			goto create_qp_exit2;
 		}
-		nr_qes = my_qp->ipz_squeue.queue_length /
+
+		my_qp->sq_map.entries = my_qp->ipz_squeue.queue_length /
 			 my_qp->ipz_squeue.qe_size;
-		my_qp->sq_map = vmalloc(nr_qes *
+		my_qp->sq_map.map = vmalloc(my_qp->sq_map.entries *
 					sizeof(struct ehca_qmap_entry));
-		if (!my_qp->sq_map) {
+		if (!my_qp->sq_map.map) {
 			ehca_err(pd->device, "Couldn't allocate squeue "
 				 "map ret=%i", ret);
 			goto create_qp_exit3;
 		}
+		INIT_LIST_HEAD(&my_qp->sq_err_node);
+		/* to avoid the generation of bogus flush CQEs */
+		reset_queue_map(&my_qp->sq_map);
 	}
 
 	if (HAS_RQ(my_qp)) {
@@ -736,6 +782,25 @@  static struct ehca_qp *internal_create_q
 				 "and pages ret=%i", ret);
 			goto create_qp_exit4;
 		}
+
+		my_qp->rq_map.entries = my_qp->ipz_rqueue.queue_length /
+			my_qp->ipz_rqueue.qe_size;
+		my_qp->rq_map.map = vmalloc(my_qp->rq_map.entries *
+				sizeof(struct ehca_qmap_entry));
+		if (!my_qp->rq_map.map) {
+			ehca_err(pd->device, "Couldn't allocate squeue "
+					"map ret=%i", ret);
+			goto create_qp_exit5;
+		}
+		INIT_LIST_HEAD(&my_qp->rq_err_node);
+		/* to avoid the generation of bogus flush CQEs */
+		reset_queue_map(&my_qp->rq_map);
+	} else if (init_attr->srq) {
+		/* this is a base QP, use the queue map of the SRQ */
+		my_qp->rq_map = my_srq->rq_map;
+		INIT_LIST_HEAD(&my_qp->rq_err_node);
+
+		my_qp->ipz_rqueue = my_srq->ipz_rqueue;
 	}
 
 	if (is_srq) {
@@ -799,7 +864,7 @@  static struct ehca_qp *internal_create_q
 		if (ret) {
 			ehca_err(pd->device,
 				 "Couldn't assign qp to send_cq ret=%i", ret);
-			goto create_qp_exit6;
+			goto create_qp_exit7;
 		}
 	}
 
@@ -825,25 +890,29 @@  static struct ehca_qp *internal_create_q
 		if (ib_copy_to_udata(udata, &resp, sizeof resp)) {
 			ehca_err(pd->device, "Copy to udata failed");
 			ret = -EINVAL;
-			goto create_qp_exit7;
+			goto create_qp_exit8;
 		}
 	}
 
 	return my_qp;
 
-create_qp_exit7:
+create_qp_exit8:
 	ehca_cq_unassign_qp(my_qp->send_cq, my_qp->real_qp_num);
 
-create_qp_exit6:
+create_qp_exit7:
 	kfree(my_qp->mod_qp_parm);
 
+create_qp_exit6:
+	if (HAS_RQ(my_qp))
+		vfree(my_qp->rq_map.map);
+
 create_qp_exit5:
 	if (HAS_RQ(my_qp))
 		ipz_queue_dtor(my_pd, &my_qp->ipz_rqueue);
 
 create_qp_exit4:
 	if (HAS_SQ(my_qp))
-		vfree(my_qp->sq_map);
+		vfree(my_qp->sq_map.map);
 
 create_qp_exit3:
 	if (HAS_SQ(my_qp))
@@ -1035,6 +1104,101 @@  static int prepare_sqe_rts(struct ehca_q
 	return 0;
 }
 
+static int calc_left_cqes(u64 wqe_p, struct ipz_queue *ipz_queue,
+			  struct ehca_queue_map *qmap)
+{
+	void *wqe_v;
+	u64 q_ofs;
+	u32 wqe_idx;
+
+	/* convert real to abs address */
+	wqe_p = wqe_p & (~(1UL << 63));
+
+	wqe_v = abs_to_virt(wqe_p);
+
+	if (ipz_queue_abs_to_offset(ipz_queue, wqe_p, &q_ofs)) {
+		ehca_gen_err("Invalid offset for calculating left cqes "
+				"wqe_p=%#lx wqe_v=%p\n", wqe_p, wqe_v);
+		return -EFAULT;
+	}
+
+	wqe_idx = q_ofs / ipz_queue->qe_size;
+	if (wqe_idx < qmap->tail)
+		qmap->left_to_poll = (qmap->entries - qmap->tail) + wqe_idx;
+	else
+		qmap->left_to_poll = wqe_idx - qmap->tail;
+
+	return 0;
+}
+
+static int check_for_left_cqes(struct ehca_qp *my_qp, struct ehca_shca *shca)
+{
+	u64 h_ret;
+	void *send_wqe_p, *recv_wqe_p;
+	int ret;
+	unsigned long flags;
+	int qp_num = my_qp->ib_qp.qp_num;
+
+	/* this hcall is not supported on base QPs */
+	if (my_qp->ext_type != EQPT_SRQBASE) {
+		/* get send and receive wqe pointer */
+		h_ret = hipz_h_disable_and_get_wqe(shca->ipz_hca_handle,
+				my_qp->ipz_qp_handle, &my_qp->pf,
+				&send_wqe_p, &recv_wqe_p, 4);
+		if (h_ret != H_SUCCESS) {
+			ehca_err(&shca->ib_device, "disable_and_get_wqe() "
+				 "failed ehca_qp=%p qp_num=%x h_ret=%li",
+				 my_qp, qp_num, h_ret);
+			return ehca2ib_return_code(h_ret);
+		}
+
+		/*
+		 * acquire lock to ensure that nobody is polling the cq which
+		 * could mean that the qmap->tail pointer is in an
+		 * inconsistent state.
+		 */
+		spin_lock_irqsave(&my_qp->send_cq->spinlock, flags);
+		ret = calc_left_cqes((u64)send_wqe_p, &my_qp->ipz_squeue,
+				&my_qp->sq_map);
+		spin_unlock_irqrestore(&my_qp->send_cq->spinlock, flags);
+		if (ret)
+			return ret;
+
+
+		spin_lock_irqsave(&my_qp->recv_cq->spinlock, flags);
+		ret = calc_left_cqes((u64)recv_wqe_p, &my_qp->ipz_rqueue,
+				&my_qp->rq_map);
+		spin_unlock_irqrestore(&my_qp->recv_cq->spinlock, flags);
+		if (ret)
+			return ret;
+	} else {
+		spin_lock_irqsave(&my_qp->send_cq->spinlock, flags);
+		my_qp->sq_map.left_to_poll = 0;
+		spin_unlock_irqrestore(&my_qp->send_cq->spinlock, flags);
+
+		spin_lock_irqsave(&my_qp->recv_cq->spinlock, flags);
+		my_qp->rq_map.left_to_poll = 0;
+		spin_unlock_irqrestore(&my_qp->recv_cq->spinlock, flags);
+	}
+
+	/* this assures flush cqes being generated only for pending wqes */
+	if ((my_qp->sq_map.left_to_poll == 0) &&
+				(my_qp->rq_map.left_to_poll == 0)) {
+		spin_lock_irqsave(&my_qp->send_cq->spinlock, flags);
+		ehca_add_to_err_list(my_qp, 1);
+		spin_unlock_irqrestore(&my_qp->send_cq->spinlock, flags);
+
+		if (HAS_RQ(my_qp)) {
+			spin_lock_irqsave(&my_qp->recv_cq->spinlock, flags);
+			ehca_add_to_err_list(my_qp, 0);
+			spin_unlock_irqrestore(&my_qp->recv_cq->spinlock,
+					flags);
+		}
+	}
+
+	return 0;
+}
+
 /*
  * internal_modify_qp with circumvention to handle aqp0 properly
  * smi_reset2init indicates if this is an internal reset-to-init-call for
@@ -1539,10 +1703,27 @@  static int internal_modify_qp(struct ib_
 			goto modify_qp_exit2;
 		}
 	}
+	if ((qp_new_state == IB_QPS_ERR) && (qp_cur_state != IB_QPS_ERR)) {
+		ret = check_for_left_cqes(my_qp, shca);
+		if (ret)
+			goto modify_qp_exit2;
+	}
 
 	if (statetrans == IB_QPST_ANY2RESET) {
 		ipz_qeit_reset(&my_qp->ipz_rqueue);
 		ipz_qeit_reset(&my_qp->ipz_squeue);
+
+		if (qp_cur_state == IB_QPS_ERR) {
+			del_from_err_list(my_qp->send_cq, &my_qp->sq_err_node);
+
+			if (HAS_RQ(my_qp))
+				del_from_err_list(my_qp->recv_cq,
+						  &my_qp->rq_err_node);
+		}
+		reset_queue_map(&my_qp->sq_map);
+
+		if (HAS_RQ(my_qp))
+			reset_queue_map(&my_qp->rq_map);
 	}
 
 	if (attr_mask & IB_QP_QKEY)
@@ -1958,6 +2139,16 @@  static int internal_destroy_qp(struct ib
 	idr_remove(&ehca_qp_idr, my_qp->token);
 	write_unlock_irqrestore(&ehca_qp_idr_lock, flags);
 
+	/*
+	 * SRQs will never get into an error list and do not have a recv_cq,
+	 * so we need to skip them here.
+	 */
+	if (HAS_RQ(my_qp) && !IS_SRQ(my_qp))
+		del_from_err_list(my_qp->recv_cq, &my_qp->rq_err_node);
+
+	if (HAS_SQ(my_qp))
+		del_from_err_list(my_qp->send_cq, &my_qp->sq_err_node);
+
 	/* now wait until all pending events have completed */
 	wait_event(my_qp->wait_completion, !atomic_read(&my_qp->nr_events));
 
@@ -1983,7 +2174,7 @@  static int internal_destroy_qp(struct ib
 	if (qp_type == IB_QPT_GSI) {
 		struct ib_event event;
 		ehca_info(dev, "device %s: port %x is inactive.",
-			  shca->ib_device.name, port_num);
+				shca->ib_device.name, port_num);
 		event.device = &shca->ib_device;
 		event.event = IB_EVENT_PORT_ERR;
 		event.element.port_num = port_num;
@@ -1991,11 +2182,15 @@  static int internal_destroy_qp(struct ib
 		ib_dispatch_event(&event);
 	}
 
-	if (HAS_RQ(my_qp))
+	if (HAS_RQ(my_qp)) {
 		ipz_queue_dtor(my_pd, &my_qp->ipz_rqueue);
+
+		vfree(my_qp->rq_map.map);
+	}
 	if (HAS_SQ(my_qp)) {
 		ipz_queue_dtor(my_pd, &my_qp->ipz_squeue);
-		vfree(my_qp->sq_map);
+
+		vfree(my_qp->sq_map.map);
 	}
 	kmem_cache_free(qp_cache, my_qp);
 	atomic_dec(&shca->num_qps);
--- infiniband.git.orig/drivers/infiniband/hw/ehca/ehca_iverbs.h
+++ infiniband.git/drivers/infiniband/hw/ehca/ehca_iverbs.h
@@ -197,6 +197,8 @@  void ehca_poll_eqs(unsigned long data);
 int ehca_calc_ipd(struct ehca_shca *shca, int port,
 		  enum ib_rate path_rate, u32 *ipd);
 
+void ehca_add_to_err_list(struct ehca_qp *qp, int on_sq);
+
 #ifdef CONFIG_PPC_64K_PAGES
 void *ehca_alloc_fw_ctrlblock(gfp_t flags);
 void ehca_free_fw_ctrlblock(void *ptr);