RDMA/iw_cxgb4: Low resource fixes for connection manager

Pre-allocate buffers for sending various control messages to close
connection, abort connection, etc so that we gracefully handle
connections when system is running out of memory.

Signed-off-by: Steve Wise <[e-mail]>
Signed-off-by: Hariprasad Shenai <[e-mail]>
Signed-off-by: Doug Ledford <[e-mail]>

drivers/infiniband/hw/cxgb4/cm.c

···
 	return;
 }
 
+static int alloc_ep_skb_list(struct sk_buff_head *ep_skb_list, int size)
+{
+	struct sk_buff *skb;
+	unsigned int i;
+	size_t len;
+
+	len = roundup(sizeof(union cpl_wr_size), 16);
+	for (i = 0; i < size; i++) {
+		skb = alloc_skb(len, GFP_KERNEL);
+		if (!skb)
+			goto fail;
+		skb_queue_tail(ep_skb_list, skb);
+	}
+	return 0;
+fail:
+	skb_queue_purge(ep_skb_list);
+	return -ENOMEM;
+}
+
 static void *alloc_ep(int size, gfp_t gfp)
 {
 	struct c4iw_ep_common *epc;
···
 		if (ep->mpa_skb)
 			kfree_skb(ep->mpa_skb);
 	}
+	if (!skb_queue_empty(&ep->com.ep_skb_list))
+		skb_queue_purge(&ep->com.ep_skb_list);
 	kfree(ep);
 }
 
···
 	}
 }
 
-static int send_flowc(struct c4iw_ep *ep, struct sk_buff *skb)
+static int send_flowc(struct c4iw_ep *ep)
 {
-	unsigned int flowclen = 80;
 	struct fw_flowc_wr *flowc;
+	struct sk_buff *skb = skb_dequeue(&ep->com.ep_skb_list);
 	int i;
 	u16 vlan = ep->l2t->vlan;
 	int nparams;
+
+	if (WARN_ON(!skb))
+		return -ENOMEM;
 
 	if (vlan == CPL_L2T_VLAN_NONE)
 		nparams = 8;
 	else
 		nparams = 9;
 
-	skb = get_skb(skb, flowclen, GFP_KERNEL);
-	flowc = (struct fw_flowc_wr *)__skb_put(skb, flowclen);
+	flowc = (struct fw_flowc_wr *)__skb_put(skb, FLOWC_LEN);
 
 	flowc->op_to_nparams = cpu_to_be32(FW_WR_OP_V(FW_FLOWC_WR) |
 					   FW_FLOWC_WR_NPARAMS_V(nparams));
-	flowc->flowid_len16 = cpu_to_be32(FW_WR_LEN16_V(DIV_ROUND_UP(flowclen,
+	flowc->flowid_len16 = cpu_to_be32(FW_WR_LEN16_V(DIV_ROUND_UP(FLOWC_LEN,
 					  16)) | FW_WR_FLOWID_V(ep->hwtid));
 
 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
···
 	return c4iw_ofld_send(&ep->com.dev->rdev, skb);
 }
 
-static int send_halfclose(struct c4iw_ep *ep, gfp_t gfp)
+static int send_halfclose(struct c4iw_ep *ep)
 {
 	struct cpl_close_con_req *req;
-	struct sk_buff *skb;
+	struct sk_buff *skb = skb_dequeue(&ep->com.ep_skb_list);
 	int wrlen = roundup(sizeof *req, 16);
 
 	PDBG("%s ep %p tid %u\n", __func__, ep, ep->hwtid);
-	skb = get_skb(NULL, wrlen, gfp);
-	if (!skb) {
-		printk(KERN_ERR MOD "%s - failed to alloc skb\n", __func__);
+	if (WARN_ON(!skb))
 		return -ENOMEM;
-	}
+
 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
 	t4_set_arp_err_handler(skb, NULL, arp_failure_discard);
 	req = (struct cpl_close_con_req *) skb_put(skb, wrlen);
···
 	return c4iw_l2t_send(&ep->com.dev->rdev, skb, ep->l2t);
 }
 
-static int send_abort(struct c4iw_ep *ep, struct sk_buff *skb, gfp_t gfp)
+static int send_abort(struct c4iw_ep *ep)
 {
 	struct cpl_abort_req *req;
 	int wrlen = roundup(sizeof *req, 16);
+	struct sk_buff *req_skb = skb_dequeue(&ep->com.ep_skb_list);
 
 	PDBG("%s ep %p tid %u\n", __func__, ep, ep->hwtid);
-	skb = get_skb(skb, wrlen, gfp);
-	if (!skb) {
-		printk(KERN_ERR MOD "%s - failed to alloc skb.\n",
-		       __func__);
+	if (WARN_ON(!req_skb))
 		return -ENOMEM;
-	}
-	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
-	t4_set_arp_err_handler(skb, ep, abort_arp_failure);
-	req = (struct cpl_abort_req *) skb_put(skb, wrlen);
+
+	set_wr_txq(req_skb, CPL_PRIORITY_DATA, ep->txq_idx);
+	t4_set_arp_err_handler(req_skb, ep, abort_arp_failure);
+	req = (struct cpl_abort_req *)skb_put(req_skb, wrlen);
 	memset(req, 0, wrlen);
 	INIT_TP_WR(req, ep->hwtid);
 	OPCODE_TID(req) = cpu_to_be32(MK_OPCODE_TID(CPL_ABORT_REQ, ep->hwtid));
 	req->cmd = CPL_ABORT_SEND_RST;
-	return c4iw_l2t_send(&ep->com.dev->rdev, skb, ep->l2t);
+	return c4iw_l2t_send(&ep->com.dev->rdev, req_skb, ep->l2t);
 }
 
 static void best_mtu(const unsigned short *mtus, unsigned short mtu,
···
 	set_bit(ACT_ESTAB, &ep->com.history);
 
 	/* start MPA negotiation */
-	ret = send_flowc(ep, NULL);
+	ret = send_flowc(ep);
 	if (ret)
 		goto err;
 	if (ep->retry_with_mpa_v1)
···
 static int c4iw_reconnect(struct c4iw_ep *ep)
 {
 	int err = 0;
+	int size = 0;
 	struct sockaddr_in *laddr = (struct sockaddr_in *)
 				    &ep->com.cm_id->m_local_addr;
 	struct sockaddr_in *raddr = (struct sockaddr_in *)
···
 	PDBG("%s qp %p cm_id %p\n", __func__, ep->com.qp, ep->com.cm_id);
 	init_timer(&ep->timer);
 	c4iw_init_wr_wait(&ep->com.wr_wait);
+
+	/* When MPA revision is different on nodes, the node with MPA_rev=2
+	 * tries to reconnect with MPA_rev 1 for the same EP through
+	 * c4iw_reconnect(), where the same EP is assigned with new tid for
+	 * further connection establishment. As we are using the same EP pointer
+	 * for reconnect, few skbs are used during the previous c4iw_connect(),
+	 * which leaves the EP with inadequate skbs for further
+	 * c4iw_reconnect(), Further causing an assert BUG_ON() due to empty
+	 * skb_list() during peer_abort(). Allocate skbs which is already used.
+	 */
+	size = (CN_MAX_CON_BUF - skb_queue_len(&ep->com.ep_skb_list));
+	if (alloc_ep_skb_list(&ep->com.ep_skb_list, size)) {
+		err = -ENOMEM;
+		goto fail1;
+	}
 
 	/*
 	 * Allocate an active TID to initiate a TCP connection.
···
 	 * response of 1st connect request.
 	 */
 	connect_reply_upcall(ep, -ECONNRESET);
+fail1:
 	c4iw_put_ep(&ep->com);
 out:
 	return err;
···
 	if (peer_mss && child_ep->mtu > (peer_mss + hdrs))
 		child_ep->mtu = peer_mss + hdrs;
 
+	skb_queue_head_init(&child_ep->com.ep_skb_list);
+	if (alloc_ep_skb_list(&child_ep->com.ep_skb_list, CN_MAX_CON_BUF))
+		goto fail;
+
 	state_set(&child_ep->com, CONNECTING);
 	child_ep->com.dev = dev;
 	child_ep->com.cm_id = NULL;
···
 			       (const u32 *)&sin6->sin6_addr.s6_addr, 1);
 	}
 	goto out;
+fail:
+	c4iw_put_ep(&child_ep->com);
 reject:
 	reject_cr(dev, hwtid, skb);
 	if (parent_ep)
···
 	ep->com.state = MPA_REQ_WAIT;
 	start_ep_timer(ep);
 	set_bit(PASS_ESTAB, &ep->com.history);
-	ret = send_flowc(ep, skb);
+	ret = send_flowc(ep);
 	mutex_unlock(&ep->com.mutex);
 	if (ret)
 		c4iw_ep_disconnect(ep, 1, GFP_KERNEL);
···
 	}
 	mutex_unlock(&ep->com.mutex);
 
-	rpl_skb = get_skb(skb, sizeof(*rpl), GFP_KERNEL);
-	if (!rpl_skb) {
-		printk(KERN_ERR MOD "%s - cannot allocate skb!\n",
-		       __func__);
+	rpl_skb = skb_dequeue(&ep->com.ep_skb_list);
+	if (WARN_ON(!rpl_skb)) {
 		release = 1;
 		goto out;
 	}
···
 		err = -ENOMEM;
 		goto out;
 	}
+
+	skb_queue_head_init(&ep->com.ep_skb_list);
+	if (alloc_ep_skb_list(&ep->com.ep_skb_list, CN_MAX_CON_BUF)) {
+		err = -ENOMEM;
+		goto fail1;
+	}
+
 	init_timer(&ep->timer);
 	ep->plen = conn_param->private_data_len;
 	if (ep->plen)
···
 	if (!ep->com.qp) {
 		PDBG("%s qpn 0x%x not found!\n", __func__, conn_param->qpn);
 		err = -EINVAL;
-		goto fail1;
+		goto fail2;
 	}
 	ref_qp(ep);
 	PDBG("%s qpn 0x%x qp %p cm_id %p\n", __func__, conn_param->qpn,
···
 	if (ep->atid == -1) {
 		printk(KERN_ERR MOD "%s - cannot alloc atid.\n", __func__);
 		err = -ENOMEM;
-		goto fail1;
+		goto fail2;
 	}
 	insert_handle(dev, &dev->atid_idr, ep, ep->atid);
 
···
 		if (raddr->sin_addr.s_addr == htonl(INADDR_ANY)) {
 			err = pick_local_ipaddrs(dev, cm_id);
 			if (err)
-				goto fail1;
+				goto fail2;
 		}
 
 		/* find a route */
···
 		if (ipv6_addr_type(&raddr6->sin6_addr) == IPV6_ADDR_ANY) {
 			err = pick_local_ip6addrs(dev, cm_id);
 			if (err)
-				goto fail1;
+				goto fail2;
 		}
 
 		/* find a route */
···
 	if (!ep->dst) {
 		printk(KERN_ERR MOD "%s - cannot find route.\n", __func__);
 		err = -EHOSTUNREACH;
-		goto fail2;
+		goto fail3;
 	}
 
 	err = import_ep(ep, iptype, ra, ep->dst, ep->com.dev, true,
 			ep->com.dev->rdev.lldi.adapter_type, cm_id->tos);
 	if (err) {
 		printk(KERN_ERR MOD "%s - cannot alloc l2e.\n", __func__);
-		goto fail3;
+		goto fail4;
 	}
 
 	PDBG("%s txq_idx %u tx_chan %u smac_idx %u rss_qid %u l2t_idx %u\n",
···
 		goto out;
 
 	cxgb4_l2t_release(ep->l2t);
-fail3:
+fail4:
 	dst_release(ep->dst);
-fail2:
+fail3:
 	remove_handle(ep->com.dev, &ep->com.dev->atid_idr, ep->atid);
 	cxgb4_free_atid(ep->com.dev->rdev.lldi.tids, ep->atid);
-fail1:
+fail2:
+	skb_queue_purge(&ep->com.ep_skb_list);
 	deref_cm_id(&ep->com);
+fail1:
 	c4iw_put_ep(&ep->com);
 out:
 	return err;
···
 		err = -ENOMEM;
 		goto fail1;
 	}
+	skb_queue_head_init(&ep->com.ep_skb_list);
 	PDBG("%s ep %p\n", __func__, ep);
 	ep->com.cm_id = cm_id;
 	ref_cm_id(&ep->com);
···
 	case MPA_REQ_RCVD:
 	case MPA_REP_SENT:
 	case FPDU_MODE:
+	case CONNECTING:
 		close = 1;
 		if (abrupt)
 			ep->com.state = ABORTING;
···
 		if (abrupt) {
 			set_bit(EP_DISC_ABORT, &ep->com.history);
 			close_complete_upcall(ep, -ECONNRESET);
-			ret = send_abort(ep, NULL, gfp);
+			ret = send_abort(ep);
 		} else {
 			set_bit(EP_DISC_CLOSE, &ep->com.history);
-			ret = send_halfclose(ep, gfp);
+			ret = send_halfclose(ep);
 		}
 		if (ret) {
 			set_bit(EP_DISC_FAIL, &ep->com.history);

drivers/infiniband/hw/cxgb4/iw_cxgb4.h

···
 	CM_ID_DEREFED		= 28,
 };
 
+enum conn_pre_alloc_buffers {
+	CN_ABORT_REQ_BUF,
+	CN_ABORT_RPL_BUF,
+	CN_CLOSE_CON_REQ_BUF,
+	CN_DESTROY_BUF,
+	CN_FLOWC_BUF,
+	CN_MAX_CON_BUF
+};
+
+#define FLOWC_LEN 80
+union cpl_wr_size {
+	struct cpl_abort_req abrt_req;
+	struct cpl_abort_rpl abrt_rpl;
+	struct fw_ri_wr ri_req;
+	struct cpl_close_con_req close_req;
+	char flowc_buf[FLOWC_LEN];
+};
+
 struct c4iw_ep_common {
 	struct iw_cm_id *cm_id;
 	struct c4iw_qp *qp;
 	struct c4iw_dev *dev;
+	struct sk_buff_head ep_skb_list;
 	enum c4iw_ep_state state;
 	struct kref kref;
 	struct mutex mutex;

drivers/infiniband/hw/cxgb4/qp.c

···
 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
 	     qhp->ep->hwtid);
 
-	skb = alloc_skb(sizeof *wqe, gfp);
-	if (!skb)
+	skb = skb_dequeue(&qhp->ep->com.ep_skb_list);
+	if (WARN_ON(!skb))
 		return;
+
 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
 
 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
···
 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
 	     ep->hwtid);
 
-	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
-	if (!skb)
+	skb = skb_dequeue(&ep->com.ep_skb_list);
+	if (WARN_ON(!skb))
 		return -ENOMEM;
+
 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
 
 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));