RDMA/cxgb3: Support peer-2-peer connection setup

Open MPI, Intel MPI and other applications don't respect the iWARP
requirement that the client (active) side of the connection send the
first RDMA message. This class of application connection setup is
called peer-to-peer. Typically once the connection is setup, _both_
sides want to send data.

This patch enables supporting peer-to-peer over the chelsio RNIC by
enforcing this iWARP requirement in the driver itself as part of RDMA
connection setup.

Connection setup is extended, when the peer2peer module option is 1,
such that the MPA initiator will send a 0B Read (the RTR) just after
connection setup. The MPA responder will suspend SQ processing until
the RTR message is received and reply-to.

In the longer term, this will be handled in a standardized way by
enhancing the MPA negotiation so peers can indicate whether they
want/need the RTR and what type of RTR (0B read, 0B write, or 0B send)
should be sent. This will be done by standardizing a few bits of the
private data in order to negotiate all this. However this patch
enables peer-to-peer applications now and allows most of the required
firmware and driver changes to be done and tested now.

Design:

- Add a module option, peer2peer, to enable this mode.

- New firmware support for peer-to-peer mode:

- a new bit in the rdma_init WR to tell it to do peer-2-peer
and what form of RTR message to send or expect.

- process _all_ preposted recvs before moving the connection
into rdma mode.

- passive side: defer completing the rdma_init WR until all
pre-posted recvs are processed. Suspend SQ processing until
the RTR is received.

- active side: expect and process the 0B read WR on offload TX
queue. Defer completing the rdma_init WR until all
pre-posted recvs are processed. Suspend SQ processing until
the 0B read WR is processed from the offload TX queue.

- If peer2peer is set, driver posts 0B read request on offload TX
queue just after posting the rdma_init WR to the offload TX queue.

- Add CQ poll logic to ignore unsolicitied read responses.

Signed-off-by: Steve Wise <[e-mail]>
Signed-off-by: Roland Dreier <[e-mail]>

drivers/infiniband/hw/cxgb3/cxio_hal.c

···
 	ptr = cq->sw_rptr;
 	while (!Q_EMPTY(ptr, cq->sw_wptr)) {
 		cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
-		if ((SQ_TYPE(*cqe) || (CQE_OPCODE(*cqe) == T3_READ_RESP)) &&
+		if ((SQ_TYPE(*cqe) ||
+		     ((CQE_OPCODE(*cqe) == T3_READ_RESP) && wq->oldest_read)) &&
 		    (CQE_QPID(*cqe) == wq->qpid))
 			(*count)++;
 		ptr++;
···
 	wqe->mpaattrs = attr->mpaattrs;
 	wqe->qpcaps = attr->qpcaps;
 	wqe->ulpdu_size = cpu_to_be16(attr->tcp_emss);
-	wqe->flags = cpu_to_be32(attr->flags);
+	wqe->rqe_count = cpu_to_be16(attr->rqe_count);
+	wqe->flags_rtr_type = cpu_to_be16(attr->flags|V_RTR_TYPE(attr->rtr_type));
 	wqe->ord = cpu_to_be32(attr->ord);
 	wqe->ird = cpu_to_be32(attr->ird);
 	wqe->qp_dma_addr = cpu_to_be64(attr->qp_dma_addr);
···
 	 *	4) cq_type is RQ_TYPE not SQ_TYPE.
 	 */
 	if (RQ_TYPE(*hw_cqe) && (CQE_OPCODE(*hw_cqe) == T3_READ_RESP)) {
+
+		/*
+		 * If this is an unsolicited read response, then the read
+		 * was generated by the kernel driver as part of peer-2-peer
+		 * connection setup.  So ignore the completion.
+		 */
+		if (!wq->oldest_read) {
+			if (CQE_STATUS(*hw_cqe))
+				wq->error = 1;
+			ret = -1;
+			goto skip_cqe;
+		}
 
 		/*
 		 * Don't write to the HWCQ, so create a new read req CQE

drivers/infiniband/hw/cxgb3/cxio_wr.h

···
 	uP_RI_QP_STAG0_ENABLE = 0x10
 } __attribute__ ((packed));
 
+enum rdma_init_rtr_types {
+	RTR_READ = 1,
+	RTR_WRITE = 2,
+	RTR_SEND = 3,
+};
+
+#define S_RTR_TYPE	2
+#define M_RTR_TYPE	0x3
+#define V_RTR_TYPE(x)	((x) << S_RTR_TYPE)
+#define G_RTR_TYPE(x)	((((x) >> S_RTR_TYPE)) & M_RTR_TYPE)
+
 struct t3_rdma_init_attr {
 	u32 tid;
 	u32 qpid;
···
 	u32 ird;
 	u64 qp_dma_addr;
 	u32 qp_dma_size;
-	u32 flags;
+	enum rdma_init_rtr_types rtr_type;
+	u16 flags;
+	u16 rqe_count;
 	u32 irs;
 };
 
···
 	u8 mpaattrs;		/* 5 */
 	u8 qpcaps;
 	__be16 ulpdu_size;
-	__be32 flags;		/* bits 31-1 - reservered */
-				/* bit     0 - set if RECV posted */
+	__be16 flags_rtr_type;
+	__be16 rqe_count;
 	__be32 ord;		/* 6 */
 	__be32 ird;
 	__be64 qp_dma_addr;	/* 7 */
···
 };
 
 enum rdma_init_wr_flags {
-	RECVS_POSTED = (1<<0),
+	MPA_INITIATOR = (1<<0),
 	PRIV_QP = (1<<1),
 };
 

drivers/infiniband/hw/cxgb3/iwch_cm.c

···
 	NULL,
 };
 
+int peer2peer = 0;
+module_param(peer2peer, int, 0644);
+MODULE_PARM_DESC(peer2peer, "Support peer2peer ULPs (default=0)");
+
 static int ep_timeout_secs = 10;
 module_param(ep_timeout_secs, int, 0644);
 MODULE_PARM_DESC(ep_timeout_secs, "CM Endpoint operation timeout "
···
 	skb_reset_transport_header(skb);
 	len = skb->len;
 	req = (struct tx_data_wr *) skb_push(skb, sizeof(*req));
-	req->wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_TX_DATA));
+	req->wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_TX_DATA)|F_WR_COMPL);
 	req->wr_lo = htonl(V_WR_TID(ep->hwtid));
 	req->len = htonl(len);
 	req->param = htonl(V_TX_PORT(ep->l2t->smt_idx) |
···
 	set_arp_failure_handler(skb, arp_failure_discard);
 	skb_reset_transport_header(skb);
 	req = (struct tx_data_wr *) skb_push(skb, sizeof(*req));
-	req->wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_TX_DATA));
+	req->wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_TX_DATA)|F_WR_COMPL);
 	req->wr_lo = htonl(V_WR_TID(ep->hwtid));
 	req->len = htonl(mpalen);
 	req->param = htonl(V_TX_PORT(ep->l2t->smt_idx) |
···
 	skb_reset_transport_header(skb);
 	len = skb->len;
 	req = (struct tx_data_wr *) skb_push(skb, sizeof(*req));
-	req->wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_TX_DATA));
+	req->wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_TX_DATA)|F_WR_COMPL);
 	req->wr_lo = htonl(V_WR_TID(ep->hwtid));
 	req->len = htonl(len);
 	req->param = htonl(V_TX_PORT(ep->l2t->smt_idx) |
···
 	 * the MPA header is valid.
 	 */
 	state_set(&ep->com, FPDU_MODE);
+	ep->mpa_attr.initiator = 1;
 	ep->mpa_attr.crc_enabled = (mpa->flags & MPA_CRC) | crc_enabled ? 1 : 0;
 	ep->mpa_attr.recv_marker_enabled = markers_enabled;
 	ep->mpa_attr.xmit_marker_enabled = mpa->flags & MPA_MARKERS ? 1 : 0;
···
 	/* bind QP and TID with INIT_WR */
 	err = iwch_modify_qp(ep->com.qp->rhp,
 			     ep->com.qp, mask, &attrs, 1);
-	if (!err)
-		goto out;
+	if (err)
+		goto err;
+
+	if (peer2peer && iwch_rqes_posted(ep->com.qp) == 0) {
+		iwch_post_zb_read(ep->com.qp);
+	}
+
+	goto out;
 err:
 	abort_connection(ep, skb, GFP_KERNEL);
 out:
···
 	 * If we get here we have accumulated the entire mpa
 	 * start reply message including private data.
 	 */
+	ep->mpa_attr.initiator = 0;
 	ep->mpa_attr.crc_enabled = (mpa->flags & MPA_CRC) | crc_enabled ? 1 : 0;
 	ep->mpa_attr.recv_marker_enabled = markers_enabled;
 	ep->mpa_attr.xmit_marker_enabled = mpa->flags & MPA_MARKERS ? 1 : 0;
···
 
 	PDBG("%s ep %p credits %u\n", __func__, ep, credits);
 
-	if (credits == 0)
+	if (credits == 0) {
+		PDBG(KERN_ERR "%s 0 credit ack  ep %p state %u\n",
+			__func__, ep, state_read(&ep->com));
 		return CPL_RET_BUF_DONE;
+	}
+
 	BUG_ON(credits != 1);
-	BUG_ON(ep->mpa_skb == NULL);
-	kfree_skb(ep->mpa_skb);
-	ep->mpa_skb = NULL;
 	dst_confirm(ep->dst);
-	if (state_read(&ep->com) == MPA_REP_SENT) {
-		ep->com.rpl_done = 1;
-		PDBG("waking up ep %p\n", ep);
-		wake_up(&ep->com.waitq);
+	if (!ep->mpa_skb) {
+		PDBG("%s rdma_init wr_ack ep %p state %u\n",
+			__func__, ep, state_read(&ep->com));
+		if (ep->mpa_attr.initiator) {
+			PDBG("%s initiator ep %p state %u\n",
+				__func__, ep, state_read(&ep->com));
+			if (peer2peer)
+				iwch_post_zb_read(ep->com.qp);
+		} else {
+			PDBG("%s responder ep %p state %u\n",
+				__func__, ep, state_read(&ep->com));
+			ep->com.rpl_done = 1;
+			wake_up(&ep->com.waitq);
+		}
+	} else {
+		PDBG("%s lsm ack ep %p state %u freeing skb\n",
+			__func__, ep, state_read(&ep->com));
+		kfree_skb(ep->mpa_skb);
+		ep->mpa_skb = NULL;
 	}
 	return CPL_RET_BUF_DONE;
 }
···
 	if (err)
 		goto err;
 
+	/* if needed, wait for wr_ack */
+	if (iwch_rqes_posted(qp)) {
+		wait_event(ep->com.waitq, ep->com.rpl_done);
+		err = ep->com.rpl_err;
+		if (err)
+			goto err;
+	}
+
 	err = send_mpa_reply(ep, conn_param->private_data,
 			     conn_param->private_data_len);
 	if (err)
 		goto err;
 
-	/* wait for wr_ack */
-	wait_event(ep->com.waitq, ep->com.rpl_done);
-	err = ep->com.rpl_err;
-	if (err)
-		goto err;
 
 	state_set(&ep->com, FPDU_MODE);
 	established_upcall(ep);

drivers/infiniband/hw/cxgb3/iwch_cm.h

···
 
 int __init iwch_cm_init(void);
 void __exit iwch_cm_term(void);
+extern int peer2peer;
 
 #endif				/* _IWCH_CM_H_ */

drivers/infiniband/hw/cxgb3/iwch_provider.h

···
 };
 
 struct iwch_mpa_attributes {
+	u8 initiator;
 	u8 recv_marker_enabled;
 	u8 xmit_marker_enabled;	/* iWARP: enable inbound Read Resp. */
 	u8 crc_enabled;
···
 	IWCH_QP_QUERY_TEST_USERWRITE = 0x32	/* Test special */
 };
 
+u16 iwch_rqes_posted(struct iwch_qp *qhp);
 int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 		      struct ib_send_wr **bad_wr);
 int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
···
 			     struct ib_mw_bind *mw_bind);
 int iwch_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
 int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg);
+int iwch_post_zb_read(struct iwch_qp *qhp);
 int iwch_register_device(struct iwch_dev *dev);
 void iwch_unregister_device(struct iwch_dev *dev);
 int iwch_quiesce_qps(struct iwch_cq *chp);

drivers/infiniband/hw/cxgb3/iwch_qp.c

···
 	}
 }
 
+int iwch_post_zb_read(struct iwch_qp *qhp)
+{
+	union t3_wr *wqe;
+	struct sk_buff *skb;
+	u8 flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
+
+	PDBG("%s enter\n", __func__);
+	skb = alloc_skb(40, GFP_KERNEL);
+	if (!skb) {
+		printk(KERN_ERR "%s cannot send zb_read!!\n", __func__);
+		return -ENOMEM;
+	}
+	wqe = (union t3_wr *)skb_put(skb, sizeof(struct t3_rdma_read_wr));
+	memset(wqe, 0, sizeof(struct t3_rdma_read_wr));
+	wqe->read.rdmaop = T3_READ_REQ;
+	wqe->read.reserved[0] = 0;
+	wqe->read.reserved[1] = 0;
+	wqe->read.reserved[2] = 0;
+	wqe->read.rem_stag = cpu_to_be32(1);
+	wqe->read.rem_to = cpu_to_be64(1);
+	wqe->read.local_stag = cpu_to_be32(1);
+	wqe->read.local_len = cpu_to_be32(0);
+	wqe->read.local_to = cpu_to_be64(1);
+	wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_READ));
+	wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid)|
+						V_FW_RIWR_LEN(flit_cnt));
+	skb->priority = CPL_PRIORITY_DATA;
+	return cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
+}
+
 /*
  * This posts a TERMINATE with layer=RDMA, type=catastrophic.
  */
···
 
 
 /*
- * Return non zero if at least one RECV was pre-posted.
+ * Return count of RECV WRs posted
  */
-static int rqes_posted(struct iwch_qp *qhp)
+u16 iwch_rqes_posted(struct iwch_qp *qhp)
 {
-	return fw_riwrh_opcode((struct fw_riwrh *)qhp->wq.queue) == T3_WR_RCV;
+	union t3_wr *wqe = qhp->wq.queue;
+	u16 count = 0;
+	while ((count+1) != 0 && fw_riwrh_opcode((struct fw_riwrh *)wqe) == T3_WR_RCV) {
+		count++;
+		wqe++;
+	}
+	PDBG("%s qhp %p count %u\n", __func__, qhp, count);
+	return count;
 }
 
 static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
···
 	init_attr.ird = qhp->attr.max_ird;
 	init_attr.qp_dma_addr = qhp->wq.dma_addr;
 	init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
-	init_attr.flags = rqes_posted(qhp) ? RECVS_POSTED : 0;
+	init_attr.rqe_count = iwch_rqes_posted(qhp);
+	init_attr.flags = qhp->attr.mpa_attr.initiator ? MPA_INITIATOR : 0;
 	init_attr.flags |= capable(CAP_NET_BIND_SERVICE) ? PRIV_QP : 0;
+	if (peer2peer) {
+		init_attr.rtr_type = RTR_READ;
+		if (init_attr.ord == 0 && qhp->attr.mpa_attr.initiator)
+			init_attr.ord = 1;
+		if (init_attr.ird == 0 && !qhp->attr.mpa_attr.initiator)
+			init_attr.ird = 1;
+	} else
+		init_attr.rtr_type = 0;
 	init_attr.irs = qhp->ep->rcv_seq;
 	PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
 	     "flags 0x%x qpcaps 0x%x\n", __func__,

drivers/net/cxgb3/version.h

···
 #define DRV_VERSION "1.0-ko"
 
 /* Firmware version */
-#define FW_VERSION_MAJOR 5
+#define FW_VERSION_MAJOR 6
 #define FW_VERSION_MINOR 0
 #define FW_VERSION_MICRO 0
 #endif				/* __CHELSIO_VERSION_H */