IB/hfi1: Implement CCA for TID RDMA protocol

Currently, FECN handling is not implemented on TID RDMA expected receive
packets and therefore CCA can't be turned on when TID RDMA is
enabled. This patch adds the CCA support to TID RDMA protocol by:

- modifying FECN RSM rule to include kernel receive contexts

- For TID_RDMA READ RESP or TID RDMA ACK packet, a CNP will be sent out if
  the FECN bit is set. For other TID RDMA packets that generate at least
  one response packet, the BECN bit will be set in the first response
  packet

- Copying expected packet data to destination buffer when FECN bit is set
  in the TID RDMA READ RESP or TID RDMA WRITE DATA packet. In this case,
  the expected packet is received as an eager packet

- Handling the TID sequence error for subsequent normal expected packets.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

drivers/infiniband/hw/hfi1/chip.c

@@ -13297,15 +13297,18 @@ static int set_up_context_variables(struct hfi1_devdata *dd)
 	/*
 	 * The RMT entries are currently allocated as shown below:
 	 * 1. QOS (0 to 128 entries);
-	 * 2. FECN for PSM (num_user_contexts + num_vnic_contexts);
+	 * 2. FECN (num_kernel_context - 1 + num_user_contexts +
+	 *    num_vnic_contexts);
 	 * 3. VNIC (num_vnic_contexts).
-	 * It should be noted that PSM FECN oversubscribe num_vnic_contexts
+	 * It should be noted that FECN oversubscribe num_vnic_contexts
 	 * entries of RMT because both VNIC and PSM could allocate any receive
 	 * context between dd->first_dyn_alloc_text and dd->num_rcv_contexts,
 	 * and PSM FECN must reserve an RMT entry for each possible PSM receive
 	 * context.
 	 */
 	rmt_count = qos_rmt_entries(dd, NULL, NULL) + (num_vnic_contexts * 2);
+	if (HFI1_CAP_IS_KSET(TID_RDMA))
+		rmt_count += num_kernel_contexts - 1;
 	if (rmt_count + n_usr_ctxts > NUM_MAP_ENTRIES) {
 		user_rmt_reduced = NUM_MAP_ENTRIES - rmt_count;
 		dd_dev_err(dd,
@@ -14288,37 +14291,43 @@ static void init_qos(struct hfi1_devdata *dd, struct rsm_map_table *rmt)
 	init_qpmap_table(dd, FIRST_KERNEL_KCTXT, dd->n_krcv_queues - 1);
 }
 
-static void init_user_fecn_handling(struct hfi1_devdata *dd,
-				    struct rsm_map_table *rmt)
+static void init_fecn_handling(struct hfi1_devdata *dd,
+			       struct rsm_map_table *rmt)
 {
 	struct rsm_rule_data rrd;
 	u64 reg;
-	int i, idx, regoff, regidx;
+	int i, idx, regoff, regidx, start;
 	u8 offset;
 	u32 total_cnt;
 
+	if (HFI1_CAP_IS_KSET(TID_RDMA))
+		/* Exclude context 0 */
+		start = 1;
+	else
+		start = dd->first_dyn_alloc_ctxt;
+
+	total_cnt = dd->num_rcv_contexts - start;
+
 	/* there needs to be enough room in the map table */
-	total_cnt = dd->num_rcv_contexts - dd->first_dyn_alloc_ctxt;
 	if (rmt->used + total_cnt >= NUM_MAP_ENTRIES) {
-		dd_dev_err(dd, "User FECN handling disabled - too many user contexts allocated\n");
+		dd_dev_err(dd, "FECN handling disabled - too many contexts allocated\n");
 		return;
 	}
 
 	/*
 	 * RSM will extract the destination context as an index into the
 	 * map table.  The destination contexts are a sequential block
-	 * in the range first_dyn_alloc_ctxt...num_rcv_contexts-1 (inclusive).
+	 * in the range start...num_rcv_contexts-1 (inclusive).
 	 * Map entries are accessed as offset + extracted value.  Adjust
 	 * the added offset so this sequence can be placed anywhere in
 	 * the table - as long as the entries themselves do not wrap.
 	 * There are only enough bits in offset for the table size, so
 	 * start with that to allow for a "negative" offset.
 	 */
-	offset = (u8)(NUM_MAP_ENTRIES + (int)rmt->used -
-						(int)dd->first_dyn_alloc_ctxt);
+	offset = (u8)(NUM_MAP_ENTRIES + rmt->used - start);
 
-	for (i = dd->first_dyn_alloc_ctxt, idx = rmt->used;
-				i < dd->num_rcv_contexts; i++, idx++) {
+	for (i = start, idx = rmt->used; i < dd->num_rcv_contexts;
+	     i++, idx++) {
 		/* replace with identity mapping */
 		regoff = (idx % 8) * 8;
 		regidx = idx / 8;
@@ -14440,7 +14449,7 @@ static void init_rxe(struct hfi1_devdata *dd)
 	rmt = alloc_rsm_map_table(dd);
 	/* set up QOS, including the QPN map table */
 	init_qos(dd, rmt);
-	init_user_fecn_handling(dd, rmt);
+	init_fecn_handling(dd, rmt);
 	complete_rsm_map_table(dd, rmt);
 	/* record number of used rsm map entries for vnic */
 	dd->vnic.rmt_start = rmt->used;

drivers/infiniband/hw/hfi1/driver.c

@@ -516,7 +516,9 @@ bool hfi1_process_ecn_slowpath(struct rvt_qp *qp, struct hfi1_packet *pkt,
 	 */
 	do_cnp = prescan ||
 		(opcode >= IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST &&
-		 opcode <= IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE);
+		 opcode <= IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE) ||
+		opcode == TID_OP(READ_RESP) ||
+		opcode == TID_OP(ACK);
 
 	/* Call appropriate CNP handler */
 	if (!ignore_fecn && do_cnp && fecn)

drivers/infiniband/hw/hfi1/tid_rdma.c

@@ -127,6 +127,14 @@ static int make_tid_rdma_ack(struct rvt_qp *qp,
 			     struct hfi1_pkt_state *ps);
 static void hfi1_do_tid_send(struct rvt_qp *qp);
 static u32 read_r_next_psn(struct hfi1_devdata *dd, u8 ctxt, u8 fidx);
+static void tid_rdma_rcv_err(struct hfi1_packet *packet,
+			     struct ib_other_headers *ohdr,
+			     struct rvt_qp *qp, u32 psn, int diff, bool fecn);
+static void update_r_next_psn_fecn(struct hfi1_packet *packet,
+				   struct hfi1_qp_priv *priv,
+				   struct hfi1_ctxtdata *rcd,
+				   struct tid_rdma_flow *flow,
+				   bool fecn);
 
 static u64 tid_rdma_opfn_encode(struct tid_rdma_params *p)
 {
@@ -2234,7 +2242,7 @@ void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet)
 	struct ib_reth *reth;
 	struct hfi1_qp_priv *qpriv = qp->priv;
 	u32 bth0, psn, len, rkey;
-	bool is_fecn;
+	bool fecn;
 	u8 next;
 	u64 vaddr;
 	int diff;
@@ -2244,7 +2252,7 @@ void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet)
 	if (hfi1_ruc_check_hdr(ibp, packet))
 		return;
 
-	is_fecn = process_ecn(qp, packet);
+	fecn = process_ecn(qp, packet);
 	psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
 	trace_hfi1_rsp_rcv_tid_read_req(qp, psn);
 
@@ -2263,9 +2271,8 @@ void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet)
 
 	diff = delta_psn(psn, qp->r_psn);
 	if (unlikely(diff)) {
-		if (tid_rdma_rcv_error(packet, ohdr, qp, psn, diff))
-			return;
-		goto send_ack;
+		tid_rdma_rcv_err(packet, ohdr, qp, psn, diff, fecn);
+		return;
 	}
 
 	/* We've verified the request, insert it into the ack queue. */
@@ -2317,11 +2324,11 @@ void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet)
 
 	/* Schedule the send tasklet. */
 	qp->s_flags |= RVT_S_RESP_PENDING;
+	if (fecn)
+		qp->s_flags |= RVT_S_ECN;
 	hfi1_schedule_send(qp);
 
 	spin_unlock_irqrestore(&qp->s_lock, flags);
-	if (is_fecn)
-		goto send_ack;
 	return;
 
 nack_inv_unlock:
@@ -2338,8 +2345,6 @@ void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet)
 	rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
 	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
 	qp->r_ack_psn = qp->r_psn;
-send_ack:
-	hfi1_send_rc_ack(packet, is_fecn);
 }
 
 u32 hfi1_build_tid_rdma_read_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,
@@ -2456,12 +2461,12 @@ void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet)
 	struct tid_rdma_request *req;
 	struct tid_rdma_flow *flow;
 	u32 opcode, aeth;
-	bool is_fecn;
+	bool fecn;
 	unsigned long flags;
 	u32 kpsn, ipsn;
 
 	trace_hfi1_sender_rcv_tid_read_resp(qp);
-	is_fecn = process_ecn(qp, packet);
+	fecn = process_ecn(qp, packet);
 	kpsn = mask_psn(be32_to_cpu(ohdr->bth[2]));
 	aeth = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.aeth);
 	opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff;
@@ -2475,9 +2480,39 @@ void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet)
 	flow = &req->flows[req->clear_tail];
 	/* When header suppression is disabled */
 	if (cmp_psn(ipsn, flow->flow_state.ib_lpsn)) {
+		update_r_next_psn_fecn(packet, priv, rcd, flow, fecn);
+
 		if (cmp_psn(kpsn, flow->flow_state.r_next_psn))
 			goto ack_done;
 		flow->flow_state.r_next_psn = mask_psn(kpsn + 1);
+		/*
+		 * Copy the payload to destination buffer if this packet is
+		 * delivered as an eager packet due to RSM rule and FECN.
+		 * The RSM rule selects FECN bit in BTH and SH bit in
+		 * KDETH header and therefore will not match the last
+		 * packet of each segment that has SH bit cleared.
+		 */
+		if (fecn && packet->etype == RHF_RCV_TYPE_EAGER) {
+			struct rvt_sge_state ss;
+			u32 len;
+			u32 tlen = packet->tlen;
+			u16 hdrsize = packet->hlen;
+			u8 pad = packet->pad;
+			u8 extra_bytes = pad + packet->extra_byte +
+				(SIZE_OF_CRC << 2);
+			u32 pmtu = qp->pmtu;
+
+			if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
+				goto ack_op_err;
+			len = restart_sge(&ss, req->e.swqe, ipsn, pmtu);
+			if (unlikely(len < pmtu))
+				goto ack_op_err;
+			rvt_copy_sge(qp, &ss, packet->payload, pmtu, false,
+				     false);
+			/* Raise the sw sequence check flag for next packet */
+			priv->s_flags |= HFI1_R_TID_SW_PSN;
+		}
+
 		goto ack_done;
 	}
 	flow->flow_state.r_next_psn = mask_psn(kpsn + 1);
@@ -2544,8 +2579,6 @@ void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet)
 
 ack_done:
 	spin_unlock_irqrestore(&qp->s_lock, flags);
-	if (is_fecn)
-		hfi1_send_rc_ack(packet, is_fecn);
 }
 
 void hfi1_kern_read_tid_flow_free(struct rvt_qp *qp)
@@ -3678,7 +3711,7 @@ void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet)
 	struct hfi1_qp_priv *qpriv = qp->priv;
 	struct tid_rdma_request *req;
 	u32 bth0, psn, len, rkey, num_segs;
-	bool is_fecn;
+	bool fecn;
 	u8 next;
 	u64 vaddr;
 	int diff;
@@ -3687,7 +3720,7 @@ void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet)
 	if (hfi1_ruc_check_hdr(ibp, packet))
 		return;
 
-	is_fecn = process_ecn(qp, packet);
+	fecn = process_ecn(qp, packet);
 	psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
 	trace_hfi1_rsp_rcv_tid_write_req(qp, psn);
 
@@ -3704,9 +3737,8 @@ void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet)
 	num_segs = DIV_ROUND_UP(len, qpriv->tid_rdma.local.max_len);
 	diff = delta_psn(psn, qp->r_psn);
 	if (unlikely(diff)) {
-		if (tid_rdma_rcv_error(packet, ohdr, qp, psn, diff))
-			return;
-		goto send_ack;
+		tid_rdma_rcv_err(packet, ohdr, qp, psn, diff, fecn);
+		return;
 	}
 
 	/*
@@ -3822,11 +3854,11 @@ void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet)
 
 	/* Schedule the send tasklet. */
 	qp->s_flags |= RVT_S_RESP_PENDING;
+	if (fecn)
+		qp->s_flags |= RVT_S_ECN;
 	hfi1_schedule_send(qp);
 
 	spin_unlock_irqrestore(&qp->s_lock, flags);
-	if (is_fecn)
-		goto send_ack;
 	return;
 
 nack_inv_unlock:
@@ -3843,8 +3875,6 @@ void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet)
 	rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
 	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
 	qp->r_ack_psn = qp->r_psn;
-send_ack:
-	hfi1_send_rc_ack(packet, is_fecn);
 }
 
 u32 hfi1_build_tid_rdma_write_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,
@@ -4061,10 +4091,10 @@ void hfi1_rc_rcv_tid_rdma_write_resp(struct hfi1_packet *packet)
 	struct tid_rdma_flow *flow;
 	enum ib_wc_status status;
 	u32 opcode, aeth, psn, flow_psn, i, tidlen = 0, pktlen;
-	bool is_fecn;
+	bool fecn;
 	unsigned long flags;
 
-	is_fecn = process_ecn(qp, packet);
+	fecn = process_ecn(qp, packet);
 	psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
 	aeth = be32_to_cpu(ohdr->u.tid_rdma.w_rsp.aeth);
 	opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff;
@@ -4204,7 +4234,6 @@ void hfi1_rc_rcv_tid_rdma_write_resp(struct hfi1_packet *packet)
 		qpriv->s_tid_cur = i;
 	}
 	qp->s_flags &= ~HFI1_S_WAIT_TID_RESP;
-
 	hfi1_schedule_tid_send(qp);
 	goto ack_done;
 
@@ -4213,9 +4242,9 @@ void hfi1_rc_rcv_tid_rdma_write_resp(struct hfi1_packet *packet)
 ack_err:
 	rvt_error_qp(qp, status);
 ack_done:
+	if (fecn)
+		qp->s_flags |= RVT_S_ECN;
 	spin_unlock_irqrestore(&qp->s_lock, flags);
-	if (is_fecn)
-		hfi1_send_rc_ack(packet, is_fecn);
 }
 
 bool hfi1_build_tid_rdma_packet(struct rvt_swqe *wqe,
@@ -4295,7 +4324,9 @@ void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet)
 	unsigned long flags;
 	u32 psn, next;
 	u8 opcode;
+	bool fecn;
 
+	fecn = process_ecn(qp, packet);
 	psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
 	opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff;
 
@@ -4308,9 +4339,53 @@ void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet)
 	req = ack_to_tid_req(e);
 	flow = &req->flows[req->clear_tail];
 	if (cmp_psn(psn, full_flow_psn(flow, flow->flow_state.lpsn))) {
+		update_r_next_psn_fecn(packet, priv, rcd, flow, fecn);
+
 		if (cmp_psn(psn, flow->flow_state.r_next_psn))
 			goto send_nak;
+
 		flow->flow_state.r_next_psn = mask_psn(psn + 1);
+		/*
+		 * Copy the payload to destination buffer if this packet is
+		 * delivered as an eager packet due to RSM rule and FECN.
+		 * The RSM rule selects FECN bit in BTH and SH bit in
+		 * KDETH header and therefore will not match the last
+		 * packet of each segment that has SH bit cleared.
+		 */
+		if (fecn && packet->etype == RHF_RCV_TYPE_EAGER) {
+			struct rvt_sge_state ss;
+			u32 len;
+			u32 tlen = packet->tlen;
+			u16 hdrsize = packet->hlen;
+			u8 pad = packet->pad;
+			u8 extra_bytes = pad + packet->extra_byte +
+				(SIZE_OF_CRC << 2);
+			u32 pmtu = qp->pmtu;
+
+			if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
+				goto send_nak;
+			len = req->comp_seg * req->seg_len;
+			len += delta_psn(psn,
+				full_flow_psn(flow, flow->flow_state.spsn)) *
+				pmtu;
+			if (unlikely(req->total_len - len < pmtu))
+				goto send_nak;
+
+			/*
+			 * The e->rdma_sge field is set when TID RDMA WRITE REQ
+			 * is first received and is never modified thereafter.
+			 */
+			ss.sge = e->rdma_sge;
+			ss.sg_list = NULL;
+			ss.num_sge = 1;
+			ss.total_len = req->total_len;
+			rvt_skip_sge(&ss, len, false);
+			rvt_copy_sge(qp, &ss, packet->payload, pmtu, false,
+				     false);
+			/* Raise the sw sequence check flag for next packet */
+			priv->r_next_psn_kdeth = mask_psn(psn + 1);
+			priv->s_flags |= HFI1_R_TID_SW_PSN;
+		}
 		goto exit;
 	}
 	flow->flow_state.r_next_psn = mask_psn(psn + 1);
@@ -4375,6 +4450,8 @@ void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet)
 	hfi1_schedule_tid_send(qp);
 exit:
 	priv->r_next_psn_kdeth = flow->flow_state.r_next_psn;
+	if (fecn)
+		qp->s_flags |= RVT_S_ECN;
 	spin_unlock_irqrestore(&qp->s_lock, flags);
 	return;
 
@@ -4476,12 +4553,11 @@ void hfi1_rc_rcv_tid_rdma_ack(struct hfi1_packet *packet)
 	struct tid_rdma_request *req;
 	struct tid_rdma_flow *flow;
 	u32 aeth, psn, req_psn, ack_psn, fspsn, resync_psn, ack_kpsn;
-	bool is_fecn;
 	unsigned long flags;
 	u16 fidx;
 
 	trace_hfi1_tid_write_sender_rcv_tid_ack(qp, 0);
-	is_fecn = process_ecn(qp, packet);
+	process_ecn(qp, packet);
 	psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
 	aeth = be32_to_cpu(ohdr->u.tid_rdma.ack.aeth);
 	req_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.ack.verbs_psn));
@@ -4835,10 +4911,10 @@ void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet)
 	struct tid_rdma_flow *flow;
 	struct tid_flow_state *fs = &qpriv->flow_state;
 	u32 psn, generation, idx, gen_next;
-	bool is_fecn;
+	bool fecn;
 	unsigned long flags;
 
-	is_fecn = process_ecn(qp, packet);
+	fecn = process_ecn(qp, packet);
 	psn = mask_psn(be32_to_cpu(ohdr->bth[2]));
 
 	generation = mask_psn(psn + 1) >> HFI1_KDETH_BTH_SEQ_SHIFT;
@@ -4929,6 +5005,8 @@ void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet)
 	qpriv->s_flags |= RVT_S_ACK_PENDING;
 	hfi1_schedule_tid_send(qp);
 bail:
+	if (fecn)
+		qp->s_flags |= RVT_S_ECN;
 	spin_unlock_irqrestore(&qp->s_lock, flags);
 }
 
@@ -5465,3 +5543,36 @@ static u32 read_r_next_psn(struct hfi1_devdata *dd, u8 ctxt, u8 fidx)
 	reg = read_uctxt_csr(dd, ctxt, RCV_TID_FLOW_TABLE + (8 * fidx));
 	return mask_psn(reg);
 }
+
+static void tid_rdma_rcv_err(struct hfi1_packet *packet,
+			     struct ib_other_headers *ohdr,
+			     struct rvt_qp *qp, u32 psn, int diff, bool fecn)
+{
+	unsigned long flags;
+
+	tid_rdma_rcv_error(packet, ohdr, qp, psn, diff);
+	if (fecn) {
+		spin_lock_irqsave(&qp->s_lock, flags);
+		qp->s_flags |= RVT_S_ECN;
+		spin_unlock_irqrestore(&qp->s_lock, flags);
+	}
+}
+
+static void update_r_next_psn_fecn(struct hfi1_packet *packet,
+				   struct hfi1_qp_priv *priv,
+				   struct hfi1_ctxtdata *rcd,
+				   struct tid_rdma_flow *flow,
+				   bool fecn)
+{
+	/*
+	 * If a start/middle packet is delivered here due to
+	 * RSM rule and FECN, we need to update the r_next_psn.
+	 */
+	if (fecn && packet->etype == RHF_RCV_TYPE_EAGER &&
+	    !(priv->s_flags & HFI1_R_TID_SW_PSN)) {
+		struct hfi1_devdata *dd = rcd->dd;
+
+		flow->flow_state.r_next_psn =
+			read_r_next_psn(dd, rcd->ctxt, flow->idx);
+	}
+}