IB/iser: Convert to CQ abstraction

Use the new CQ abstraction to simplify completions in the iSER
initiator.

Signed-off-by: Sagi Grimberg <sagig@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

1 files changed, 57 insertions, 201 deletions

diff --git a/drivers/infiniband/ulp/iser/iser_verbs.c b/drivers/infiniband/ulp/iser/iser_verbs.c
index 67c4e341291d..87a50eb5e56c 100644
--- a/drivers/infiniband/ulp/iser/iser_verbs.c
+++ b/drivers/infiniband/ulp/iser/iser_verbs.c
@@ -44,17 +44,6 @@
 #define ISER_MAX_CQ_LEN		(ISER_MAX_RX_LEN + ISER_MAX_TX_LEN + \
 				 ISCSI_ISER_MAX_CONN)
 
-static int iser_cq_poll_limit = 512;
-
-static void iser_cq_tasklet_fn(unsigned long data);
-static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
-
-static void iser_cq_event_callback(struct ib_event *cause, void *context)
-{
-	iser_err("cq event %s (%d)\n",
-		 ib_event_msg(cause->event), cause->event);
-}
-
 static void iser_qp_event_callback(struct ib_event *cause, void *context)
 {
 	iser_err("qp event %s (%d)\n",
@@ -110,27 +99,14 @@ static int iser_create_device_ib_res(struct iser_device *device)
 		goto pd_err;
 
 	for (i = 0; i < device->comps_used; i++) {
-		struct ib_cq_init_attr cq_attr = {};
 		struct iser_comp *comp = &device->comps[i];
 
-		comp->device = device;
-		cq_attr.cqe = max_cqe;
-		cq_attr.comp_vector = i;
-		comp->cq = ib_create_cq(device->ib_device,
-					iser_cq_callback,
-					iser_cq_event_callback,
-					(void *)comp,
-					&cq_attr);
+		comp->cq = ib_alloc_cq(device->ib_device, comp, max_cqe, i,
+				IB_POLL_SOFTIRQ);
 		if (IS_ERR(comp->cq)) {
 			comp->cq = NULL;
 			goto cq_err;
 		}
-
-		if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
-			goto cq_err;
-
-		tasklet_init(&comp->tasklet, iser_cq_tasklet_fn,
-			     (unsigned long)comp);
 	}
 
 	if (!iser_always_reg) {
@@ -140,7 +116,7 @@ static int iser_create_device_ib_res(struct iser_device *device)
 
 		device->mr = ib_get_dma_mr(device->pd, access);
 		if (IS_ERR(device->mr))
-			goto dma_mr_err;
+			goto cq_err;
 	}
 
 	INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
@@ -153,15 +129,12 @@ static int iser_create_device_ib_res(struct iser_device *device)
 handler_err:
 	if (device->mr)
 		ib_dereg_mr(device->mr);
-dma_mr_err:
-	for (i = 0; i < device->comps_used; i++)
-		tasklet_kill(&device->comps[i].tasklet);
 cq_err:
 	for (i = 0; i < device->comps_used; i++) {
 		struct iser_comp *comp = &device->comps[i];
 
 		if (comp->cq)
-			ib_destroy_cq(comp->cq);
+			ib_free_cq(comp->cq);
 	}
 	ib_dealloc_pd(device->pd);
 pd_err:
@@ -182,8 +155,7 @@ static void iser_free_device_ib_res(struct iser_device *device)
 	for (i = 0; i < device->comps_used; i++) {
 		struct iser_comp *comp = &device->comps[i];
 
-		tasklet_kill(&comp->tasklet);
-		ib_destroy_cq(comp->cq);
+		ib_free_cq(comp->cq);
 		comp->cq = NULL;
 	}
 
@@ -723,13 +695,13 @@ int iser_conn_terminate(struct iser_conn *iser_conn)
 				 iser_conn, err);
 
 		/* post an indication that all flush errors were consumed */
-		err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
+		err = ib_post_send(ib_conn->qp, &ib_conn->last, &bad_wr);
 		if (err) {
-			iser_err("conn %p failed to post beacon", ib_conn);
+			iser_err("conn %p failed to post last wr", ib_conn);
 			return 1;
 		}
 
-		wait_for_completion(&ib_conn->flush_comp);
+		wait_for_completion(&ib_conn->last_comp);
 	}
 
 	return 1;
@@ -966,14 +938,21 @@ static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *eve
 
 void iser_conn_init(struct iser_conn *iser_conn)
 {
+	struct ib_conn *ib_conn = &iser_conn->ib_conn;
+
 	iser_conn->state = ISER_CONN_INIT;
-	iser_conn->ib_conn.post_recv_buf_count = 0;
-	init_completion(&iser_conn->ib_conn.flush_comp);
 	init_completion(&iser_conn->stop_completion);
 	init_completion(&iser_conn->ib_completion);
 	init_completion(&iser_conn->up_completion);
 	INIT_LIST_HEAD(&iser_conn->conn_list);
 	mutex_init(&iser_conn->state_mutex);
+
+	ib_conn->post_recv_buf_count = 0;
+	ib_conn->reg_cqe.done = iser_reg_comp;
+	ib_conn->last_cqe.done = iser_last_comp;
+	ib_conn->last.wr_cqe = &ib_conn->last_cqe;
+	ib_conn->last.opcode = IB_WR_SEND;
+	init_completion(&ib_conn->last_comp);
 }
 
  /**
@@ -999,9 +978,6 @@ int iser_connect(struct iser_conn   *iser_conn,
 
 	iser_conn->state = ISER_CONN_PENDING;
 
-	ib_conn->beacon.wr_id = ISER_BEACON_WRID;
-	ib_conn->beacon.opcode = IB_WR_SEND;
-
 	ib_conn->cma_id = rdma_create_id(&init_net, iser_cma_handler,
 					 (void *)iser_conn,
 					 RDMA_PS_TCP, IB_QPT_RC);
@@ -1044,56 +1020,60 @@ connect_failure:
 
 int iser_post_recvl(struct iser_conn *iser_conn)
 {
-	struct ib_recv_wr rx_wr, *rx_wr_failed;
 	struct ib_conn *ib_conn = &iser_conn->ib_conn;
 	struct iser_login_desc *desc = &iser_conn->login_desc;
+	struct ib_recv_wr wr, *wr_failed;
 	int ib_ret;
 
 	desc->sge.addr = desc->rsp_dma;
 	desc->sge.length = ISER_RX_LOGIN_SIZE;
 	desc->sge.lkey = ib_conn->device->pd->local_dma_lkey;
 
-	rx_wr.wr_id = (uintptr_t)desc;
-	rx_wr.sg_list = &desc->sge;
-	rx_wr.num_sge = 1;
-	rx_wr.next = NULL;
+	desc->cqe.done = iser_login_rsp;
+	wr.wr_cqe = &desc->cqe;
+	wr.sg_list = &desc->sge;
+	wr.num_sge = 1;
+	wr.next = NULL;
 
 	ib_conn->post_recv_buf_count++;
-	ib_ret	= ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
+	ib_ret = ib_post_recv(ib_conn->qp, &wr, &wr_failed);
 	if (ib_ret) {
 		iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 		ib_conn->post_recv_buf_count--;
 	}
+
 	return ib_ret;
 }
 
 int iser_post_recvm(struct iser_conn *iser_conn, int count)
 {
-	struct ib_recv_wr *rx_wr, *rx_wr_failed;
-	int i, ib_ret;
 	struct ib_conn *ib_conn = &iser_conn->ib_conn;
 	unsigned int my_rx_head = iser_conn->rx_desc_head;
 	struct iser_rx_desc *rx_desc;
+	struct ib_recv_wr *wr, *wr_failed;
+	int i, ib_ret;
 
-	for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
-		rx_desc		= &iser_conn->rx_descs[my_rx_head];
-		rx_wr->wr_id	= (uintptr_t)rx_desc;
-		rx_wr->sg_list	= &rx_desc->rx_sg;
-		rx_wr->num_sge	= 1;
-		rx_wr->next	= rx_wr + 1;
+	for (wr = ib_conn->rx_wr, i = 0; i < count; i++, wr++) {
+		rx_desc = &iser_conn->rx_descs[my_rx_head];
+		rx_desc->cqe.done = iser_task_rsp;
+		wr->wr_cqe = &rx_desc->cqe;
+		wr->sg_list = &rx_desc->rx_sg;
+		wr->num_sge = 1;
+		wr->next = wr + 1;
 		my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
 	}
 
-	rx_wr--;
-	rx_wr->next = NULL; /* mark end of work requests list */
+	wr--;
+	wr->next = NULL; /* mark end of work requests list */
 
 	ib_conn->post_recv_buf_count += count;
-	ib_ret	= ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
+	ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &wr_failed);
 	if (ib_ret) {
 		iser_err("ib_post_recv failed ret=%d\n", ib_ret);
 		ib_conn->post_recv_buf_count -= count;
 	} else
 		iser_conn->rx_desc_head = my_rx_head;
+
 	return ib_ret;
 }
 
@@ -1114,7 +1094,7 @@ int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
 				      DMA_TO_DEVICE);
 
 	wr->next = NULL;
-	wr->wr_id = (uintptr_t)tx_desc;
+	wr->wr_cqe = &tx_desc->cqe;
 	wr->sg_list = tx_desc->tx_sg;
 	wr->num_sge = tx_desc->num_sge;
 	wr->opcode = IB_WR_SEND;
@@ -1128,148 +1108,6 @@ int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
 	return ib_ret;
 }
 
-/**
- * is_iser_tx_desc - Indicate if the completion wr_id
- *     is a TX descriptor or not.
- * @iser_conn: iser connection
- * @wr_id: completion WR identifier
- *
- * Since we cannot rely on wc opcode in FLUSH errors
- * we must work around it by checking if the wr_id address
- * falls in the iser connection rx_descs buffer. If so
- * it is an RX descriptor, otherwize it is a TX.
- */
-static inline bool
-is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
-{
-	void *start = iser_conn->rx_descs;
-	int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
-
-	if (wr_id >= start && wr_id < start + len)
-		return false;
-
-	return true;
-}
-
-/**
- * iser_handle_comp_error() - Handle error completion
- * @ib_conn:   connection RDMA resources
- * @wc:        work completion
- *
- * Notes: We may handle a FLUSH error completion and in this case
- *        we only cleanup in case TX type was DATAOUT. For non-FLUSH
- *        error completion we should also notify iscsi layer that
- *        connection is failed (in case we passed bind stage).
- */
-static void
-iser_handle_comp_error(struct ib_conn *ib_conn,
-		       struct ib_wc *wc)
-{
-	struct iser_conn *iser_conn = to_iser_conn(ib_conn);
-	void *wr_id = (void *)(uintptr_t)wc->wr_id;
-
-	if (wc->status != IB_WC_WR_FLUSH_ERR)
-		if (iser_conn->iscsi_conn)
-			iscsi_conn_failure(iser_conn->iscsi_conn,
-					   ISCSI_ERR_CONN_FAILED);
-
-	if (wc->wr_id == ISER_FASTREG_LI_WRID)
-		return;
-
-	if (is_iser_tx_desc(iser_conn, wr_id)) {
-		struct iser_tx_desc *desc = wr_id;
-
-		if (desc->type == ISCSI_TX_DATAOUT)
-			kmem_cache_free(ig.desc_cache, desc);
-	} else {
-		ib_conn->post_recv_buf_count--;
-	}
-}
-
-/**
- * iser_handle_wc - handle a single work completion
- * @wc: work completion
- *
- * Soft-IRQ context, work completion can be either
- * SEND or RECV, and can turn out successful or
- * with error (or flush error).
- */
-static void iser_handle_wc(struct ib_wc *wc)
-{
-	struct ib_conn *ib_conn;
-	struct iser_tx_desc *tx_desc;
-	struct iser_rx_desc *rx_desc;
-
-	ib_conn = wc->qp->qp_context;
-	if (likely(wc->status == IB_WC_SUCCESS)) {
-		if (wc->opcode == IB_WC_RECV) {
-			rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
-			iser_rcv_completion(rx_desc, wc->byte_len,
-					    ib_conn);
-		} else
-		if (wc->opcode == IB_WC_SEND) {
-			tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
-			iser_snd_completion(tx_desc, ib_conn);
-		} else {
-			iser_err("Unknown wc opcode %d\n", wc->opcode);
-		}
-	} else {
-		if (wc->status != IB_WC_WR_FLUSH_ERR)
-			iser_err("%s (%d): wr id %llx vend_err %x\n",
-				 ib_wc_status_msg(wc->status), wc->status,
-				 wc->wr_id, wc->vendor_err);
-		else
-			iser_dbg("%s (%d): wr id %llx\n",
-				 ib_wc_status_msg(wc->status), wc->status,
-				 wc->wr_id);
-
-		if (wc->wr_id == ISER_BEACON_WRID)
-			/* all flush errors were consumed */
-			complete(&ib_conn->flush_comp);
-		else
-			iser_handle_comp_error(ib_conn, wc);
-	}
-}
-
-/**
- * iser_cq_tasklet_fn - iSER completion polling loop
- * @data: iSER completion context
- *
- * Soft-IRQ context, polling connection CQ until
- * either CQ was empty or we exausted polling budget
- */
-static void iser_cq_tasklet_fn(unsigned long data)
-{
-	struct iser_comp *comp = (struct iser_comp *)data;
-	struct ib_cq *cq = comp->cq;
-	struct ib_wc *const wcs = comp->wcs;
-	int i, n, completed = 0;
-
-	while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
-		for (i = 0; i < n; i++)
-			iser_handle_wc(&wcs[i]);
-
-		completed += n;
-		if (completed >= iser_cq_poll_limit)
-			break;
-	}
-
-	/*
-	 * It is assumed here that arming CQ only once its empty
-	 * would not cause interrupts to be missed.
-	 */
-	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
-
-	iser_dbg("got %d completions\n", completed);
-}
-
-static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
-{
-	struct iser_comp *comp = cq_context;
-
-	tasklet_schedule(&comp->tasklet);
-}
-
 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
 			     enum iser_data_dir cmd_dir, sector_t *sector)
 {
@@ -1317,3 +1155,21 @@ err:
 	/* Not alot we can do here, return ambiguous guard error */
 	return 0x1;
 }
+
+void iser_err_comp(struct ib_wc *wc, const char *type)
+{
+	if (wc->status != IB_WC_WR_FLUSH_ERR) {
+		struct iser_conn *iser_conn = to_iser_conn(wc->qp->qp_context);
+
+		iser_err("%s failure: %s (%d) vend_err %x\n", type,
+			 ib_wc_status_msg(wc->status), wc->status,
+			 wc->vendor_err);
+
+		if (iser_conn->iscsi_conn)
+			iscsi_conn_failure(iser_conn->iscsi_conn,
+					   ISCSI_ERR_CONN_FAILED);
+	} else {
+		iser_dbg("%s failure: %s (%d)\n", type,
+			 ib_wc_status_msg(wc->status), wc->status);
+	}
+}