IB/hfi1: Add 16B UD support

Add 16B bypass packet support for UD traffic types.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Signed-off-by: NDon Hiatt <don.hiatt@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

drivers/infiniband/hw/hfi1/driver.c

@@ -437,23 +437,33 @@ void hfi1_process_ecn_slowpath(struct rvt_qp *qp, struct hfi1_packet *pkt,
 			       bool do_cnp)
 {
 	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
-	struct ib_header *hdr = pkt->hdr;
 	struct ib_other_headers *ohdr = pkt->ohdr;
 	struct ib_grh *grh = pkt->grh;
 	u32 rqpn = 0, bth1;
-	u16 rlid, dlid = ib_get_dlid(hdr);
-	u8 sc, svc_type;
+	u16 pkey, rlid, dlid = ib_get_dlid(pkt->hdr);
+	u8 hdr_type, sc, svc_type;
 	bool is_mcast = false;
 
+	if (pkt->etype == RHF_RCV_TYPE_BYPASS) {
+		is_mcast = hfi1_is_16B_mcast(dlid);
+		pkey = hfi1_16B_get_pkey(pkt->hdr);
+		sc = hfi1_16B_get_sc(pkt->hdr);
+		hdr_type = HFI1_PKT_TYPE_16B;
+	} else {
+		is_mcast = (dlid > be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
+			   (dlid != be16_to_cpu(IB_LID_PERMISSIVE));
+		pkey = ib_bth_get_pkey(ohdr);
+		sc = hfi1_9B_get_sc5(pkt->hdr, pkt->rhf);
+		hdr_type = HFI1_PKT_TYPE_9B;
+	}
+
 	switch (qp->ibqp.qp_type) {
 	case IB_QPT_SMI:
 	case IB_QPT_GSI:
 	case IB_QPT_UD:
-		rlid = ib_get_slid(hdr);
-		rqpn = ib_get_sqpn(ohdr);
+		rlid = ib_get_slid(pkt->hdr);
+		rqpn = ib_get_sqpn(pkt->ohdr);
 		svc_type = IB_CC_SVCTYPE_UD;
-		is_mcast = (dlid > be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
-			(dlid != be16_to_cpu(IB_LID_PERMISSIVE));
 		break;
 	case IB_QPT_UC:
 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
@@ -469,14 +479,11 @@ void hfi1_process_ecn_slowpath(struct rvt_qp *qp, struct hfi1_packet *pkt,
 		return;
 	}
 
-	sc = hfi1_9B_get_sc5(hdr, pkt->rhf);
-
 	bth1 = be32_to_cpu(ohdr->bth[1]);
-	if (do_cnp && (bth1 & IB_FECN_SMASK)) {
-		u16 pkey = ib_bth_get_pkey(ohdr);
-
-		return_cnp(ibp, qp, rqpn, pkey, dlid, rlid, sc, grh);
-	}
+	/* Call appropriate CNP handler */
+	if (do_cnp && (bth1 & IB_FECN_SMASK))
+		hfi1_handle_cnp_tbl[hdr_type](ibp, qp, rqpn, pkey,
+					      dlid, rlid, sc, grh);
 
 	if (!is_mcast && (bth1 & IB_BECN_SMASK)) {
 		struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);

drivers/infiniband/hw/hfi1/hfi.h

@@ -831,6 +831,10 @@ struct hfi1_pportdata {
 typedef int (*rhf_rcv_function_ptr)(struct hfi1_packet *packet);
 
 typedef void (*opcode_handler)(struct hfi1_packet *packet);
+typedef void (*hfi1_make_req)(struct rvt_qp *qp,
+			      struct hfi1_pkt_state *ps,
+			      struct rvt_swqe *wqe);
+
 
 /* return values for the RHF receive functions */
 #define RHF_RCV_CONTINUE  0	/* keep going */
@@ -1373,6 +1377,13 @@ void hfi1_set_vnic_msix_info(struct hfi1_ctxtdata *rcd);
 void hfi1_reset_vnic_msix_info(struct hfi1_ctxtdata *rcd);
 
 extern const struct pci_device_id hfi1_pci_tbl[];
+void hfi1_make_ud_req_9B(struct rvt_qp *qp,
+			 struct hfi1_pkt_state *ps,
+			 struct rvt_swqe *wqe);
+
+void hfi1_make_ud_req_16B(struct rvt_qp *qp,
+			  struct hfi1_pkt_state *ps,
+			  struct rvt_swqe *wqe);
 
 /* receive packet handler dispositions */
 #define RCV_PKT_OK      0x0 /* keep going */
@@ -1507,6 +1518,18 @@ void process_becn(struct hfi1_pportdata *ppd, u8 sl,  u16 rlid, u32 lqpn,
 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn,
 		u32 pkey, u32 slid, u32 dlid, u8 sc5,
 		const struct ib_grh *old_grh);
+void return_cnp_16B(struct hfi1_ibport *ibp, struct rvt_qp *qp,
+		    u32 remote_qpn, u32 pkey, u32 slid, u32 dlid,
+		    u8 sc5, const struct ib_grh *old_grh);
+typedef void (*hfi1_handle_cnp)(struct hfi1_ibport *ibp, struct rvt_qp *qp,
+				u32 remote_qpn, u32 pkey, u32 slid, u32 dlid,
+				u8 sc5, const struct ib_grh *old_grh);
+
+/* We support only two types - 9B and 16B for now */
+static const hfi1_handle_cnp hfi1_handle_cnp_tbl[2] = {
+	[HFI1_PKT_TYPE_9B] = &return_cnp,
+	[HFI1_PKT_TYPE_16B] = &return_cnp_16B
+};
 #define PKEY_CHECK_INVALID -1
 int egress_pkey_check(struct hfi1_pportdata *ppd, __be16 *lrh, __be32 *bth,
 		      u8 sc5, int8_t s_pkey_index);
@@ -1747,12 +1770,22 @@ static inline bool process_ecn(struct rvt_qp *qp, struct hfi1_packet *pkt,
 			       bool do_cnp)
 {
 	struct ib_other_headers *ohdr = pkt->ohdr;
-	u32 bth1;
 
-	bth1 = be32_to_cpu(ohdr->bth[1]);
-	if (unlikely(bth1 & (IB_BECN_SMASK | IB_FECN_SMASK))) {
+	u32 bth1;
+	bool becn = false;
+	bool fecn = false;
+
+	if (pkt->etype == RHF_RCV_TYPE_BYPASS) {
+		fecn = hfi1_16B_get_fecn(pkt->hdr);
+		becn = hfi1_16B_get_becn(pkt->hdr);
+	} else {
+		bth1 = be32_to_cpu(ohdr->bth[1]);
+		fecn = bth1 & IB_FECN_SMASK;
+		becn = bth1 & IB_BECN_SMASK;
+	}
+	if (unlikely(fecn || becn)) {
 		hfi1_process_ecn_slowpath(qp, pkt, do_cnp);
-		return !!(bth1 & IB_FECN_SMASK);
+		return fecn;
 	}
 	return false;
 }
@@ -2315,4 +2348,80 @@ static inline bool hfi1_get_hdr_type(u32 lid, struct rdma_ah_attr *attr)
 
 	return hfi1_get_packet_type(lid);
 }
+
+static inline void hfi1_make_ext_grh(struct hfi1_packet *packet,
+				     struct ib_grh *grh, u32 slid,
+				     u32 dlid)
+{
+	struct hfi1_ibport *ibp = &packet->rcd->ppd->ibport_data;
+	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
+
+	if (!ibp)
+		return;
+
+	grh->hop_limit = 1;
+	grh->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
+	if (slid == opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B))
+		grh->sgid.global.interface_id =
+			OPA_MAKE_ID(be32_to_cpu(OPA_LID_PERMISSIVE));
+	else
+		grh->sgid.global.interface_id = OPA_MAKE_ID(slid);
+
+	/*
+	 * Upper layers (like mad) may compare the dgid in the
+	 * wc that is obtained here with the sgid_index in
+	 * the wr. Since sgid_index in wr is always 0 for
+	 * extended lids, set the dgid here to the default
+	 * IB gid.
+	 */
+	grh->dgid.global.subnet_prefix = ibp->rvp.gid_prefix;
+	grh->dgid.global.interface_id =
+		cpu_to_be64(ppd->guids[HFI1_PORT_GUID_INDEX]);
+}
+
+static inline int hfi1_get_16b_padding(u32 hdr_size, u32 payload)
+{
+	return -(hdr_size + payload + (SIZE_OF_CRC << 2) +
+		     SIZE_OF_LT) & 0x7;
+}
+
+static inline void hfi1_make_ib_hdr(struct ib_header *hdr,
+				    u16 lrh0, u16 len,
+				    u16 dlid, u16 slid)
+{
+	hdr->lrh[0] = cpu_to_be16(lrh0);
+	hdr->lrh[1] = cpu_to_be16(dlid);
+	hdr->lrh[2] = cpu_to_be16(len);
+	hdr->lrh[3] = cpu_to_be16(slid);
+}
+
+static inline void hfi1_make_16b_hdr(struct hfi1_16b_header *hdr,
+				     u32 slid, u32 dlid,
+				     u16 len, u16 pkey,
+				     u8 becn, u8 fecn, u8 l4,
+				     u8 sc)
+{
+	u32 lrh0 = 0;
+	u32 lrh1 = 0x40000000;
+	u32 lrh2 = 0;
+	u32 lrh3 = 0;
+
+	lrh0 = (lrh0 & ~OPA_16B_BECN_MASK) | (becn << OPA_16B_BECN_SHIFT);
+	lrh0 = (lrh0 & ~OPA_16B_LEN_MASK) | (len << OPA_16B_LEN_SHIFT);
+	lrh0 = (lrh0 & ~OPA_16B_LID_MASK)  | (slid & OPA_16B_LID_MASK);
+	lrh1 = (lrh1 & ~OPA_16B_FECN_MASK) | (fecn << OPA_16B_FECN_SHIFT);
+	lrh1 = (lrh1 & ~OPA_16B_SC_MASK) | (sc << OPA_16B_SC_SHIFT);
+	lrh1 = (lrh1 & ~OPA_16B_LID_MASK) | (dlid & OPA_16B_LID_MASK);
+	lrh2 = (lrh2 & ~OPA_16B_SLID_MASK) |
+		((slid >> OPA_16B_SLID_SHIFT) << OPA_16B_SLID_HIGH_SHIFT);
+	lrh2 = (lrh2 & ~OPA_16B_DLID_MASK) |
+		((dlid >> OPA_16B_DLID_SHIFT) << OPA_16B_DLID_HIGH_SHIFT);
+	lrh2 = (lrh2 & ~OPA_16B_PKEY_MASK) | (pkey << OPA_16B_PKEY_SHIFT);
+	lrh2 = (lrh2 & ~OPA_16B_L4_MASK) | l4;
+
+	hdr->lrh[0] = lrh0;
+	hdr->lrh[1] = lrh1;
+	hdr->lrh[2] = lrh2;
+	hdr->lrh[3] = lrh3;
+}
 #endif                          /* _HFI1_KERNEL_H */

drivers/infiniband/hw/hfi1/mad.c

@@ -373,12 +373,10 @@ static struct trap_node *create_trap_node(u8 type, __be16 trap_num, u32 lid)
  * Send a bad P_Key trap (ch. 14.3.8).
  */
 void hfi1_bad_pkey(struct hfi1_ibport *ibp, u32 key, u32 sl,
-		   u32 qp1, u32 qp2, u16 lid1, u16 lid2)
+		   u32 qp1, u32 qp2, u32 lid1, u32 lid2)
 {
 	struct trap_node *trap;
 	u32 lid = ppd_from_ibp(ibp)->lid;
-	u32 _lid1 = lid1;
-	u32 _lid2 = lid2;
 
 	ibp->rvp.n_pkt_drops++;
 	ibp->rvp.pkey_violations++;
@@ -389,8 +387,8 @@ void hfi1_bad_pkey(struct hfi1_ibport *ibp, u32 key, u32 sl,
 		return;
 
 	/* Send violation trap */
-	trap->data.ntc_257_258.lid1 = cpu_to_be32(_lid1);
-	trap->data.ntc_257_258.lid2 = cpu_to_be32(_lid2);
+	trap->data.ntc_257_258.lid1 = cpu_to_be32(lid1);
+	trap->data.ntc_257_258.lid2 = cpu_to_be32(lid2);
 	trap->data.ntc_257_258.key = cpu_to_be32(key);
 	trap->data.ntc_257_258.sl = sl << 3;
 	trap->data.ntc_257_258.qp1 = cpu_to_be32(qp1);

drivers/infiniband/hw/hfi1/ruc.c

@@ -649,7 +649,7 @@ static void ruc_loopback(struct rvt_qp *sqp)
  * @ibp: a pointer to the IB port
  * @hdr: a pointer to the GRH header being constructed
  * @grh: the global route address to send to
- * @hwords: the number of 32 bit words of header being sent
+ * @hwords: size of header after grh being sent in dwords
  * @nwords: the number of 32 bit words of data being sent
  *
  * Return the size of the header in 32 bit words.
@@ -661,7 +661,7 @@ u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
-	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
+	hdr->paylen = cpu_to_be16((hwords + nwords) << 2);
 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
 	hdr->next_hdr = IB_GRH_NEXT_HDR;
 	hdr->hop_limit = grh->hop_limit;

drivers/infiniband/hw/hfi1/ud.c

@@ -53,6 +53,12 @@
 #include "verbs_txreq.h"
 #include "qp.h"
 
+/* We support only two types - 9B and 16B for now */
+static const hfi1_make_req hfi1_make_ud_req_tbl[2] = {
+	[HFI1_PKT_TYPE_9B] = &hfi1_make_ud_req_9B,
+	[HFI1_PKT_TYPE_16B] = &hfi1_make_ud_req_16B
+};
+
 /**
  * ud_loopback - handle send on loopback QPs
  * @sqp: the sending QP
@@ -67,6 +73,7 @@ static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
 {
 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
 	struct hfi1_pportdata *ppd;
+	struct hfi1_qp_priv *priv = sqp->priv;
 	struct rvt_qp *qp;
 	struct rdma_ah_attr *ah_attr;
 	unsigned long flags;
@@ -102,7 +109,7 @@ static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
 
 	if (qp->ibqp.qp_num > 1) {
 		u16 pkey;
-		u16 slid;
+		u32 slid;
 		u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
 
 		pkey = hfi1_get_pkey(ibp, sqp->s_pkey_index);
@@ -176,9 +183,33 @@ static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
 
 	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
 		struct ib_grh grh;
-		const struct ib_global_route *grd = rdma_ah_read_grh(ah_attr);
+		struct ib_global_route grd = *(rdma_ah_read_grh(ah_attr));
+
+		/*
+		 * For loopback packets with extended LIDs, the
+		 * sgid_index in the GRH is 0 and the dgid is
+		 * OPA GID of the sender. While creating a response
+		 * to the loopback packet, IB core creates the new
+		 * sgid_index from the DGID and that will be the
+		 * OPA_GID_INDEX. The new dgid is from the sgid
+		 * index and that will be in the IB GID format.
+		 *
+		 * We now have a case where the sent packet had a
+		 * different sgid_index and dgid compared to the
+		 * one that was received in response.
+		 *
+		 * Fix this inconsistency.
+		 */
+		if (priv->hdr_type == HFI1_PKT_TYPE_16B) {
+			if (grd.sgid_index == 0)
+				grd.sgid_index = OPA_GID_INDEX;
 
-		hfi1_make_grh(ibp, &grh, grd, 0, 0);
+			if (ib_is_opa_gid(&grd.dgid))
+				grd.dgid.global.interface_id =
+				cpu_to_be64(ppd->guids[HFI1_PORT_GUID_INDEX]);
+		}
+
+		hfi1_make_grh(ibp, &grh, &grd, 0, 0);
 		hfi1_copy_sge(&qp->r_sge, &grh,
 			      sizeof(grh), true, false);
 		wc.wc_flags |= IB_WC_GRH;
@@ -235,7 +266,7 @@ static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
 		wc.pkey_index = 0;
 	}
 	wc.slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
-			      ((1 << ppd->lmc) - 1));
+				   ((1 << ppd->lmc) - 1));
 	/* Check for loopback when the port lid is not set */
 	if (wc.slid == 0 && sqp->ibqp.qp_type == IB_QPT_GSI)
 		wc.slid = be16_to_cpu(IB_LID_PERMISSIVE);
@@ -252,6 +283,183 @@ static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
 	rcu_read_unlock();
 }
 
+static void hfi1_make_bth_deth(struct rvt_qp *qp, struct rvt_swqe *wqe,
+			       struct ib_other_headers *ohdr,
+			       u16 *pkey, u32 extra_bytes, bool bypass)
+{
+	u32 bth0;
+	struct hfi1_ibport *ibp;
+
+	ibp = to_iport(qp->ibqp.device, qp->port_num);
+	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
+		ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
+		bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
+	} else {
+		bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
+	}
+
+	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
+		bth0 |= IB_BTH_SOLICITED;
+	bth0 |= extra_bytes << 20;
+	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI)
+		*pkey = hfi1_get_pkey(ibp, wqe->ud_wr.pkey_index);
+	else
+		*pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
+	if (!bypass)
+		bth0 |= *pkey;
+	ohdr->bth[0] = cpu_to_be32(bth0);
+	ohdr->bth[1] = cpu_to_be32(wqe->ud_wr.remote_qpn);
+	ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn));
+	/*
+	 * Qkeys with the high order bit set mean use the
+	 * qkey from the QP context instead of the WR (see 10.2.5).
+	 */
+	ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->ud_wr.remote_qkey < 0 ?
+					 qp->qkey : wqe->ud_wr.remote_qkey);
+	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
+}
+
+void hfi1_make_ud_req_9B(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
+			 struct rvt_swqe *wqe)
+{
+	u32 nwords, extra_bytes;
+	u16 len, slid, dlid, pkey;
+	u16 lrh0 = 0;
+	u8 sc5;
+	struct hfi1_qp_priv *priv = qp->priv;
+	struct ib_other_headers *ohdr;
+	struct rdma_ah_attr *ah_attr;
+	struct hfi1_pportdata *ppd;
+	struct hfi1_ibport *ibp;
+	struct ib_grh *grh;
+
+	ibp = to_iport(qp->ibqp.device, qp->port_num);
+	ppd = ppd_from_ibp(ibp);
+	ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
+
+	extra_bytes = -wqe->length & 3;
+	nwords = ((wqe->length + extra_bytes) >> 2) + SIZE_OF_CRC;
+	/* header size in dwords LRH+BTH+DETH = (8+12+8)/4. */
+	qp->s_hdrwords = 7;
+	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
+		qp->s_hdrwords++;
+
+	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
+		grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh;
+		qp->s_hdrwords += hfi1_make_grh(ibp, grh,
+						rdma_ah_read_grh(ah_attr),
+						qp->s_hdrwords - 2, nwords);
+		lrh0 = HFI1_LRH_GRH;
+		ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
+	} else {
+		lrh0 = HFI1_LRH_BTH;
+		ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
+	}
+
+	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
+	lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4;
+	if (qp->ibqp.qp_type == IB_QPT_SMI) {
+		lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
+		priv->s_sc = 0xf;
+	} else {
+		lrh0 |= (sc5 & 0xf) << 12;
+		priv->s_sc = sc5;
+	}
+
+	dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 9B);
+	if (dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
+		slid = be16_to_cpu(IB_LID_PERMISSIVE);
+	} else {
+		u16 lid = (u16)ppd->lid;
+
+		if (lid) {
+			lid |= rdma_ah_get_path_bits(ah_attr) &
+				((1 << ppd->lmc) - 1);
+			slid = lid;
+		} else {
+			slid = be16_to_cpu(IB_LID_PERMISSIVE);
+		}
+	}
+	hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, false);
+	len = qp->s_hdrwords + nwords;
+
+	/* Setup the packet */
+	ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_9B;
+	hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh,
+			 lrh0, len, dlid, slid);
+}
+
+void hfi1_make_ud_req_16B(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
+			  struct rvt_swqe *wqe)
+{
+	struct hfi1_qp_priv *priv = qp->priv;
+	struct ib_other_headers *ohdr;
+	struct rdma_ah_attr *ah_attr;
+	struct hfi1_pportdata *ppd;
+	struct hfi1_ibport *ibp;
+	u32 dlid, slid, nwords, extra_bytes;
+	u16 len, pkey;
+	u8 l4, sc5;
+
+	ibp = to_iport(qp->ibqp.device, qp->port_num);
+	ppd = ppd_from_ibp(ibp);
+	ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
+	/* header size in dwords 16B LRH+BTH+DETH = (16+12+8)/4. */
+	qp->s_hdrwords = 9;
+	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
+		qp->s_hdrwords++;
+
+	/* SW provides space for CRC and LT for bypass packets. */
+	extra_bytes = hfi1_get_16b_padding((qp->s_hdrwords << 2),
+					   wqe->length);
+	nwords = ((wqe->length + extra_bytes + SIZE_OF_LT) >> 2) + SIZE_OF_CRC;
+
+	if ((rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) &&
+	    hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) {
+		struct ib_grh *grh;
+		struct ib_global_route *grd = rdma_ah_retrieve_grh(ah_attr);
+		/*
+		 * Ensure OPA GIDs are transformed to IB gids
+		 * before creating the GRH.
+		 */
+		if (grd->sgid_index == OPA_GID_INDEX) {
+			dd_dev_warn(ppd->dd, "Bad sgid_index. sgid_index: %d\n",
+				    grd->sgid_index);
+			grd->sgid_index = 0;
+		}
+		grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh;
+		qp->s_hdrwords += hfi1_make_grh(ibp, grh, grd,
+					qp->s_hdrwords - 4, nwords);
+		ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
+		l4 = OPA_16B_L4_IB_GLOBAL;
+	} else {
+		ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
+		l4 = OPA_16B_L4_IB_LOCAL;
+	}
+
+	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
+	if (qp->ibqp.qp_type == IB_QPT_SMI)
+		priv->s_sc = 0xf;
+	else
+		priv->s_sc = sc5;
+
+	dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 16B);
+	if (!ppd->lid)
+		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
+	else
+		slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
+			   ((1 << ppd->lmc) - 1));
+
+	hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, true);
+	/* Convert dwords to flits */
+	len = (qp->s_hdrwords + nwords) >> 1;
+
+	/* Setup the packet */
+	ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_16B;
+	hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah,
+			  slid, dlid, len, pkey, 0, 0, l4, priv->s_sc);
+}
+
 /**
  * hfi1_make_ud_req - construct a UD request packet
  * @qp: the QP
@@ -263,18 +471,12 @@ static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
 int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 {
 	struct hfi1_qp_priv *priv = qp->priv;
-	struct ib_other_headers *ohdr;
 	struct rdma_ah_attr *ah_attr;
 	struct hfi1_pportdata *ppd;
 	struct hfi1_ibport *ibp;
 	struct rvt_swqe *wqe;
-	u32 nwords;
-	u32 extra_bytes;
-	u32 bth0;
-	u16 lrh0;
-	u16 lid;
 	int next_cur;
-	u8 sc5;
+	u32 lid;
 
 	ps->s_txreq = get_txreq(ps->dev, qp);
 	if (IS_ERR(ps->s_txreq))
@@ -311,13 +513,14 @@ int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 	ibp = to_iport(qp->ibqp.device, qp->port_num);
 	ppd = ppd_from_ibp(ibp);
 	ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
-	if (rdma_ah_get_dlid(ah_attr) < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
-	    rdma_ah_get_dlid(ah_attr) == be16_to_cpu(IB_LID_PERMISSIVE)) {
+	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, ah_attr);
+	if ((!hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) ||
+	    (rdma_ah_get_dlid(ah_attr) == be32_to_cpu(OPA_LID_PERMISSIVE))) {
 		lid = rdma_ah_get_dlid(ah_attr) & ~((1 << ppd->lmc) - 1);
 		if (unlikely(!loopback &&
-			     (lid == ppd->lid ||
-			      (lid == be16_to_cpu(IB_LID_PERMISSIVE) &&
-			      qp->ibqp.qp_type == IB_QPT_GSI)))) {
+			     ((lid == ppd->lid) ||
+			      ((lid == be32_to_cpu(OPA_LID_PERMISSIVE)) &&
+			       (qp->ibqp.qp_type == IB_QPT_GSI))))) {
 			unsigned long tflags = ps->flags;
 			/*
 			 * If DMAs are in progress, we can't generate
@@ -341,11 +544,6 @@ int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 	}
 
 	qp->s_cur = next_cur;
-	extra_bytes = -wqe->length & 3;
-	nwords = (wqe->length + extra_bytes) >> 2;
-
-	/* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
-	qp->s_hdrwords = 7;
 	ps->s_txreq->s_cur_size = wqe->length;
 	ps->s_txreq->ss = &qp->s_sge;
 	qp->s_srate = rdma_ah_get_static_rate(ah_attr);
@@ -356,78 +554,12 @@ int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 	qp->s_sge.num_sge = wqe->wr.num_sge;
 	qp->s_sge.total_len = wqe->length;
 
-	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
-		/* Header size in 32-bit words. */
-		qp->s_hdrwords +=
-			hfi1_make_grh(ibp,
-				      &ps->s_txreq->phdr.hdr.ibh.u.l.grh,
-				      rdma_ah_read_grh(ah_attr),
-				      qp->s_hdrwords, nwords);
-		lrh0 = HFI1_LRH_GRH;
-		ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
-		/*
-		 * Don't worry about sending to locally attached multicast
-		 * QPs.  It is unspecified by the spec. what happens.
-		 */
-	} else {
-		/* Header size in 32-bit words. */
-		lrh0 = HFI1_LRH_BTH;
-		ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
-	}
-	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
-		qp->s_hdrwords++;
-		ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
-		bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
-	} else {
-		bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
-	}
-	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
-	lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4;
-	if (qp->ibqp.qp_type == IB_QPT_SMI) {
-		lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
-		priv->s_sc = 0xf;
-	} else {
-		lrh0 |= (sc5 & 0xf) << 12;
-		priv->s_sc = sc5;
-	}
+	/* Make the appropriate header */
+	hfi1_make_ud_req_tbl[priv->hdr_type](qp, ps, qp->s_wqe);
 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
 	ps->s_txreq->sde = priv->s_sde;
 	priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
 	ps->s_txreq->psc = priv->s_sendcontext;
-	ps->s_txreq->phdr.hdr.ibh.lrh[0] = cpu_to_be16(lrh0);
-	ps->s_txreq->phdr.hdr.ibh.lrh[1] =
-		cpu_to_be16(rdma_ah_get_dlid(ah_attr));
-	ps->s_txreq->phdr.hdr.ibh.lrh[2] =
-		cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
-	if (rdma_ah_get_dlid(ah_attr) == be16_to_cpu(IB_LID_PERMISSIVE)) {
-		ps->s_txreq->phdr.hdr.ibh.lrh[3] = IB_LID_PERMISSIVE;
-	} else {
-		lid = ppd->lid;
-		if (lid) {
-			lid |= rdma_ah_get_path_bits(ah_attr) &
-				((1 << ppd->lmc) - 1);
-			ps->s_txreq->phdr.hdr.ibh.lrh[3] = cpu_to_be16(lid);
-		} else {
-			ps->s_txreq->phdr.hdr.ibh.lrh[3] = IB_LID_PERMISSIVE;
-		}
-	}
-	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
-		bth0 |= IB_BTH_SOLICITED;
-	bth0 |= extra_bytes << 20;
-	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI)
-		bth0 |= hfi1_get_pkey(ibp, wqe->ud_wr.pkey_index);
-	else
-		bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
-	ohdr->bth[0] = cpu_to_be32(bth0);
-	ohdr->bth[1] = cpu_to_be32(wqe->ud_wr.remote_qpn);
-	ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn));
-	/*
-	 * Qkeys with the high order bit set mean use the
-	 * qkey from the QP context instead of the WR (see 10.2.5).
-	 */
-	ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->ud_wr.remote_qkey < 0 ?
-					 qp->qkey : wqe->ud_wr.remote_qkey);
-	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
 	/* disarm any ahg */
 	priv->s_ahg->ahgcount = 0;
 	priv->s_ahg->ahgidx = 0;
@@ -497,6 +629,64 @@ int hfi1_lookup_pkey_idx(struct hfi1_ibport *ibp, u16 pkey)
 	return -1;
 }
 
+void return_cnp_16B(struct hfi1_ibport *ibp, struct rvt_qp *qp,
+		    u32 remote_qpn, u32 pkey, u32 slid, u32 dlid,
+		    u8 sc5, const struct ib_grh *old_grh)
+{
+	u64 pbc, pbc_flags = 0;
+	u32 bth0, plen, vl, hwords = 7;
+	u16 len;
+	u8 l4;
+	struct hfi1_16b_header hdr;
+	struct ib_other_headers *ohdr;
+	struct pio_buf *pbuf;
+	struct send_context *ctxt = qp_to_send_context(qp, sc5);
+	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
+	u32 nwords;
+
+	/* Populate length */
+	nwords = ((hfi1_get_16b_padding(hwords << 2, 0) +
+		   SIZE_OF_LT) >> 2) + SIZE_OF_CRC;
+	if (old_grh) {
+		struct ib_grh *grh = &hdr.u.l.grh;
+
+		grh->version_tclass_flow = old_grh->version_tclass_flow;
+		grh->paylen = cpu_to_be16((hwords - 4 + nwords) << 2);
+		grh->hop_limit = 0xff;
+		grh->sgid = old_grh->dgid;
+		grh->dgid = old_grh->sgid;
+		ohdr = &hdr.u.l.oth;
+		l4 = OPA_16B_L4_IB_GLOBAL;
+		hwords += sizeof(struct ib_grh) / sizeof(u32);
+	} else {
+		ohdr = &hdr.u.oth;
+		l4 = OPA_16B_L4_IB_LOCAL;
+	}
+
+	/* BIT 16 to 19 is TVER. Bit 20 to 22 is pad cnt */
+	bth0 = (IB_OPCODE_CNP << 24) | (1 << 16) |
+	       (hfi1_get_16b_padding(hwords << 2, 0) << 20);
+	ohdr->bth[0] = cpu_to_be32(bth0);
+
+	ohdr->bth[1] = cpu_to_be32(remote_qpn);
+	ohdr->bth[2] = 0; /* PSN 0 */
+
+	/* Convert dwords to flits */
+	len = (hwords + nwords) >> 1;
+	hfi1_make_16b_hdr(&hdr, slid, dlid, len, pkey, 1, 0, l4, sc5);
+
+	plen = 2 /* PBC */ + hwords + nwords;
+	pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
+	vl = sc_to_vlt(ppd->dd, sc5);
+	pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
+	if (ctxt) {
+		pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
+		if (pbuf)
+			ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
+						 &hdr, hwords);
+	}
+}
+
 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn,
 		u32 pkey, u32 slid, u32 dlid, u8 sc5,
 		const struct ib_grh *old_grh)
@@ -535,11 +725,7 @@ void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn,
 	ohdr->bth[1] = cpu_to_be32(remote_qpn | (1 << IB_BECN_SHIFT));
 	ohdr->bth[2] = 0; /* PSN 0 */
 
-	hdr.lrh[0] = cpu_to_be16(lrh0);
-	hdr.lrh[1] = cpu_to_be16(dlid);
-	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
-	hdr.lrh[3] = cpu_to_be16(slid);
-
+	hfi1_make_ib_hdr(&hdr, lrh0, hwords + SIZE_OF_CRC, dlid, slid);
 	plen = 2 /* PBC */ + hwords;
 	pbc_flags |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
 	vl = sc_to_vlt(ppd->dd, sc5);
@@ -672,18 +858,33 @@ void hfi1_ud_rcv(struct hfi1_packet *packet)
 	void *data = packet->payload;
 	u32 tlen = packet->tlen;
 	struct rvt_qp *qp = packet->qp;
-	u8 sc5 = hfi1_9B_get_sc5(hdr, packet->rhf);
+	u8 sc5 = packet->sc;
 	u8 sl_from_sc;
-	u8 extra_bytes = packet->pad;
 	u8 opcode = packet->opcode;
 	u8 sl = packet->sl;
 	u32 dlid = packet->dlid;
 	u32 slid = packet->slid;
+	u8 extra_bytes;
+	bool dlid_is_permissive;
+	bool slid_is_permissive;
 
+	extra_bytes = packet->pad + packet->extra_byte + (SIZE_OF_CRC << 2);
 	qkey = ib_get_qkey(ohdr);
 	src_qp = ib_get_sqpn(ohdr);
-	pkey = ib_bth_get_pkey(ohdr);
-	extra_bytes += (SIZE_OF_CRC << 2);
+
+	if (packet->etype == RHF_RCV_TYPE_BYPASS) {
+		u32 permissive_lid =
+			opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B);
+
+		pkey = hfi1_16B_get_pkey(packet->hdr);
+		dlid_is_permissive = (dlid == permissive_lid);
+		slid_is_permissive = (slid == permissive_lid);
+	} else {
+		hdr = packet->hdr;
+		pkey = ib_bth_get_pkey(ohdr);
+		dlid_is_permissive = (dlid == be16_to_cpu(IB_LID_PERMISSIVE));
+		slid_is_permissive = (slid == be16_to_cpu(IB_LID_PERMISSIVE));
+	}
 	sl_from_sc = ibp->sc_to_sl[sc5];
 
 	process_ecn(qp, packet, (opcode != IB_OPCODE_CNP));
@@ -701,8 +902,7 @@ void hfi1_ud_rcv(struct hfi1_packet *packet)
 	 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
 	 */
 	if (qp->ibqp.qp_num) {
-		if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
-			     hdr->lrh[3] == IB_LID_PERMISSIVE))
+		if (unlikely(dlid_is_permissive || slid_is_permissive))
 			goto drop;
 		if (qp->ibqp.qp_num > 1) {
 			if (unlikely(rcv_pkey_check(ppd, pkey, sc5, slid))) {
@@ -740,8 +940,7 @@ void hfi1_ud_rcv(struct hfi1_packet *packet)
 
 		if (tlen > 2048)
 			goto drop;
-		if ((hdr->lrh[1] == IB_LID_PERMISSIVE ||
-		     hdr->lrh[3] == IB_LID_PERMISSIVE) &&
+		if ((dlid_is_permissive || slid_is_permissive) &&
 		    smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 			goto drop;
 
@@ -794,7 +993,18 @@ void hfi1_ud_rcv(struct hfi1_packet *packet)
 		goto drop;
 	}
 	if (packet->grh) {
-		hfi1_copy_sge(&qp->r_sge, &hdr->u.l.grh,
+		hfi1_copy_sge(&qp->r_sge, packet->grh,
+			      sizeof(struct ib_grh), true, false);
+		wc.wc_flags |= IB_WC_GRH;
+	} else if (packet->etype == RHF_RCV_TYPE_BYPASS) {
+		struct ib_grh grh;
+		/*
+		 * Assuming we only created 16B on the send side
+		 * if we want to use large LIDs, since GRH was stripped
+		 * out when creating 16B, add back the GRH here.
+		 */
+		hfi1_make_ext_grh(packet, &grh, slid, dlid);
+		hfi1_copy_sge(&qp->r_sge, &grh,
 			      sizeof(struct ib_grh), true, false);
 		wc.wc_flags |= IB_WC_GRH;
 	} else {
@@ -827,14 +1037,15 @@ void hfi1_ud_rcv(struct hfi1_packet *packet)
 	} else {
 		wc.pkey_index = 0;
 	}
-
+	if (slid_is_permissive)
+		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
 	wc.slid = slid;
 	wc.sl = sl_from_sc;
 
 	/*
 	 * Save the LMC lower bits if the destination LID is a unicast LID.
 	 */
-	wc.dlid_path_bits = dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE) ? 0 :
+	wc.dlid_path_bits = hfi1_check_mcast(dlid) ? 0 :
 		dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
 	wc.port_num = qp->port_num;
 	/* Signal completion event if the solicited bit is set. */

drivers/infiniband/hw/hfi1/verbs.h

@@ -259,7 +259,7 @@ static inline int hfi1_send_ok(struct rvt_qp *qp)
  * This must be called with s_lock held.
  */
 void hfi1_bad_pkey(struct hfi1_ibport *ibp, u32 key, u32 sl,
-		   u32 qp1, u32 qp2, u16 lid1, u16 lid2);
+		   u32 qp1, u32 qp2, u32 lid1, u32 lid2);
 void hfi1_cap_mask_chg(struct rvt_dev_info *rdi, u8 port_num);
 void hfi1_sys_guid_chg(struct hfi1_ibport *ibp);
 void hfi1_node_desc_chg(struct hfi1_ibport *ibp);

include/rdma/opa_addr.h

@@ -54,6 +54,7 @@
 #define OPA_MAKE_ID(x)          (cpu_to_be64(OPA_SPECIAL_OUI << 40 | (x)))
 #define OPA_TO_IB_UCAST_LID(x) (((x) >= be16_to_cpu(IB_MULTICAST_LID_BASE)) \
 				? 0 : x)
+#define OPA_GID_INDEX		0x1
 /**
  * 0xF8 - 4 bits of multicast range and 1 bit for collective range
  * Example: For 24 bit LID space,