RDMA/cxgb3: Fixes for zero STag

Handling the zero STag in receive work request requires some extra
logic in the driver:

 - Only set the QP_PRIV bit for kernel mode QPs.

- Add a zero STag build function for recv wrs. The uP needs a PBL
  allocated and passed down in the recv WR so it can construct a HW
  PBL for the zero STag S/G entries.  Note: we need to place a few
  restrictions on zero STag usage because of this:

  1) all SGEs in a recv WR must either be zero STag or not.  No mixing.

  2) an individual SGE length cannot exceed 128MB for a zero-stag SGE.
     This should be OK since it's not really practical to allocate
     such a large chunk of pinned contiguous DMA mapped memory.

- Add an optimized non-zero-STag recv wr format for kernel users.
  This is needed to optimize both zero and non-zero STag cracking in
  the recv path for kernel users.

 - Remove the iwch_ prefix from the static build functions.

 - Bump required FW version.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>

4 files changed, 130 insertions, 26 deletions

diff --git a/drivers/infiniband/hw/cxgb3/cxio_hal.c b/drivers/infiniband/hw/cxgb3/cxio_hal.c
index 340e4181c761..f6d5747153a5 100644
--- a/drivers/infiniband/hw/cxgb3/cxio_hal.c
+++ b/drivers/infiniband/hw/cxgb3/cxio_hal.c
@@ -278,7 +278,7 @@ int cxio_create_qp(struct cxio_rdev *rdev_p, u32 kernel_domain,
 	if (!wq->qpid)
 		return -ENOMEM;
 
-	wq->rq = kzalloc(depth * sizeof(u64), GFP_KERNEL);
+	wq->rq = kzalloc(depth * sizeof(struct t3_swrq), GFP_KERNEL);
 	if (!wq->rq)
 		goto err1;
 
@@ -302,6 +302,7 @@ int cxio_create_qp(struct cxio_rdev *rdev_p, u32 kernel_domain,
 	if (!kernel_domain)
 		wq->udb = (u64)rdev_p->rnic_info.udbell_physbase +
 					(wq->qpid << rdev_p->qpshift);
+	wq->rdev = rdev_p;
 	PDBG("%s qpid 0x%x doorbell 0x%p udb 0x%llx\n", __func__,
 	     wq->qpid, wq->doorbell, (unsigned long long) wq->udb);
 	return 0;
@@ -1266,13 +1267,16 @@ proc_cqe:
 		wq->sq_rptr = CQE_WRID_SQ_WPTR(*hw_cqe);
 		PDBG("%s completing sq idx %ld\n", __func__,
 		     Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2));
-		*cookie = (wq->sq +
-			   Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2))->wr_id;
+		*cookie = wq->sq[Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2)].wr_id;
 		wq->sq_rptr++;
 	} else {
 		PDBG("%s completing rq idx %ld\n", __func__,
 		     Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
-		*cookie = *(wq->rq + Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
+		*cookie = wq->rq[Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2)].wr_id;
+		if (wq->rq[Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2)].pbl_addr)
+			cxio_hal_pblpool_free(wq->rdev,
+				wq->rq[Q_PTR2IDX(wq->rq_rptr,
+				wq->rq_size_log2)].pbl_addr, T3_STAG0_PBL_SIZE);
 		wq->rq_rptr++;
 	}
 
diff --git a/drivers/infiniband/hw/cxgb3/cxio_wr.h b/drivers/infiniband/hw/cxgb3/cxio_wr.h
index de760e9f1cc6..04618f7bfbb3 100644
--- a/drivers/infiniband/hw/cxgb3/cxio_wr.h
+++ b/drivers/infiniband/hw/cxgb3/cxio_wr.h
@@ -39,6 +39,9 @@
 
 #define T3_MAX_SGE      4
 #define T3_MAX_INLINE	64
+#define T3_STAG0_PBL_SIZE (2 * T3_MAX_SGE << 3)
+#define T3_STAG0_MAX_PBE_LEN (128 * 1024 * 1024)
+#define T3_STAG0_PAGE_SHIFT 15
 
 #define Q_EMPTY(rptr,wptr) ((rptr)==(wptr))
 #define Q_FULL(rptr,wptr,size_log2)  ( (((wptr)-(rptr))>>(size_log2)) && \
@@ -665,6 +668,11 @@ struct t3_swsq {
 	int			signaled;
 };
 
+struct t3_swrq {
+	__u64			wr_id;
+	__u32			pbl_addr;
+};
+
 /*
  * A T3 WQ implements both the SQ and RQ.
  */
@@ -681,14 +689,15 @@ struct t3_wq {
 	u32 sq_wptr;			/* sq_wptr - sq_rptr == count of */
 	u32 sq_rptr;			/* pending wrs */
 	u32 sq_size_log2;		/* sq size */
-	u64 *rq;			/* SW RQ (holds consumer wr_ids */
+	struct t3_swrq *rq;		/* SW RQ (holds consumer wr_ids */
 	u32 rq_wptr;			/* rq_wptr - rq_rptr == count of */
 	u32 rq_rptr;			/* pending wrs */
-	u64 *rq_oldest_wr;		/* oldest wr on the SW RQ */
+	struct t3_swrq *rq_oldest_wr;	/* oldest wr on the SW RQ */
 	u32 rq_size_log2;		/* rq size */
 	u32 rq_addr;			/* rq adapter address */
 	void __iomem *doorbell;		/* kernel db */
 	u64 udb;			/* user db if any */
+	struct cxio_rdev *rdev;
 };
 
 struct t3_cq {
diff --git a/drivers/infiniband/hw/cxgb3/iwch_provider.c b/drivers/infiniband/hw/cxgb3/iwch_provider.c
index 7ecfd4d638c5..b89640aa6e10 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_provider.c
+++ b/drivers/infiniband/hw/cxgb3/iwch_provider.c
@@ -1007,10 +1007,10 @@ static struct ib_qp *iwch_create_qp(struct ib_pd *pd,
 	qhp->ibqp.qp_num = qhp->wq.qpid;
 	init_timer(&(qhp->timer));
 	PDBG("%s sq_num_entries %d, rq_num_entries %d "
-	     "qpid 0x%0x qhp %p dma_addr 0x%llx size %d\n",
+	     "qpid 0x%0x qhp %p dma_addr 0x%llx size %d rq_addr 0x%x\n",
 	     __func__, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
 	     qhp->wq.qpid, qhp, (unsigned long long) qhp->wq.dma_addr,
-	     1 << qhp->wq.size_log2);
+	     1 << qhp->wq.size_log2, qhp->wq.rq_addr);
 	return &qhp->ibqp;
 }
 
diff --git a/drivers/infiniband/hw/cxgb3/iwch_qp.c b/drivers/infiniband/hw/cxgb3/iwch_qp.c
index 3b44300a3036..9a3be3a9d5dc 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_qp.c
+++ b/drivers/infiniband/hw/cxgb3/iwch_qp.c
@@ -33,10 +33,11 @@
 #include "iwch.h"
 #include "iwch_cm.h"
 #include "cxio_hal.h"
+#include "cxio_resource.h"
 
 #define NO_SUPPORT -1
 
-static int iwch_build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
 				u8 * flit_cnt)
 {
 	int i;
@@ -81,7 +82,7 @@ static int iwch_build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
 	return 0;
 }
 
-static int iwch_build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
 				 u8 *flit_cnt)
 {
 	int i;
@@ -122,7 +123,7 @@ static int iwch_build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
 	return 0;
 }
 
-static int iwch_build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
 				u8 *flit_cnt)
 {
 	if (wr->num_sge > 1)
@@ -143,7 +144,7 @@ static int iwch_build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
 	return 0;
 }
 
-static int iwch_build_fastreg(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_fastreg(union t3_wr *wqe, struct ib_send_wr *wr,
 				u8 *flit_cnt, int *wr_cnt, struct t3_wq *wq)
 {
 	int i;
@@ -185,7 +186,7 @@ static int iwch_build_fastreg(union t3_wr *wqe, struct ib_send_wr *wr,
 	return 0;
 }
 
-static int iwch_build_inv_stag(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_inv_stag(union t3_wr *wqe, struct ib_send_wr *wr,
 				u8 *flit_cnt)
 {
 	wqe->local_inv.stag = cpu_to_be32(wr->ex.invalidate_rkey);
@@ -244,23 +245,106 @@ static int iwch_sgl2pbl_map(struct iwch_dev *rhp, struct ib_sge *sg_list,
 	return 0;
 }
 
-static int iwch_build_rdma_recv(struct iwch_dev *rhp, union t3_wr *wqe,
+static int build_rdma_recv(struct iwch_qp *qhp, union t3_wr *wqe,
 				struct ib_recv_wr *wr)
 {
-	int i;
-	if (wr->num_sge > T3_MAX_SGE)
-		return -EINVAL;
+	int i, err = 0;
+	u32 pbl_addr[T3_MAX_SGE];
+	u8 page_size[T3_MAX_SGE];
+
+	err = iwch_sgl2pbl_map(qhp->rhp, wr->sg_list, wr->num_sge, pbl_addr,
+			       page_size);
+	if (err)
+		return err;
+	wqe->recv.pagesz[0] = page_size[0];
+	wqe->recv.pagesz[1] = page_size[1];
+	wqe->recv.pagesz[2] = page_size[2];
+	wqe->recv.pagesz[3] = page_size[3];
 	wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
 	for (i = 0; i < wr->num_sge; i++) {
 		wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
 		wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
+
+		/* to in the WQE == the offset into the page */
+		wqe->recv.sgl[i].to = cpu_to_be64(((u32) wr->sg_list[i].addr) %
+				(1UL << (12 + page_size[i])));
+
+		/* pbl_addr is the adapters address in the PBL */
+		wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_addr[i]);
+	}
+	for (; i < T3_MAX_SGE; i++) {
+		wqe->recv.sgl[i].stag = 0;
+		wqe->recv.sgl[i].len = 0;
+		wqe->recv.sgl[i].to = 0;
+		wqe->recv.pbl_addr[i] = 0;
+	}
+	qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
+			     qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
+	qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
+			     qhp->wq.rq_size_log2)].pbl_addr = 0;
+	return 0;
+}
+
+static int build_zero_stag_recv(struct iwch_qp *qhp, union t3_wr *wqe,
+				struct ib_recv_wr *wr)
+{
+	int i;
+	u32 pbl_addr;
+	u32 pbl_offset;
+
+
+	/*
+	 * The T3 HW requires the PBL in the HW recv descriptor to reference
+	 * a PBL entry.  So we allocate the max needed PBL memory here and pass
+	 * it to the uP in the recv WR.  The uP will build the PBL and setup
+	 * the HW recv descriptor.
+	 */
+	pbl_addr = cxio_hal_pblpool_alloc(&qhp->rhp->rdev, T3_STAG0_PBL_SIZE);
+	if (!pbl_addr)
+		return -ENOMEM;
+
+	/*
+	 * Compute the 8B aligned offset.
+	 */
+	pbl_offset = (pbl_addr - qhp->rhp->rdev.rnic_info.pbl_base) >> 3;
+
+	wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
+
+	for (i = 0; i < wr->num_sge; i++) {
+
+		/*
+		 * Use a 128MB page size. This and an imposed 128MB
+		 * sge length limit allows us to require only a 2-entry HW
+		 * PBL for each SGE.  This restriction is acceptable since
+		 * since it is not possible to allocate 128MB of contiguous
+		 * DMA coherent memory!
+		 */
+		if (wr->sg_list[i].length > T3_STAG0_MAX_PBE_LEN)
+			return -EINVAL;
+		wqe->recv.pagesz[i] = T3_STAG0_PAGE_SHIFT;
+
+		/*
+		 * T3 restricts a recv to all zero-stag or all non-zero-stag.
+		 */
+		if (wr->sg_list[i].lkey != 0)
+			return -EINVAL;
+		wqe->recv.sgl[i].stag = 0;
+		wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
 		wqe->recv.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
+		wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_offset);
+		pbl_offset += 2;
 	}
 	for (; i < T3_MAX_SGE; i++) {
+		wqe->recv.pagesz[i] = 0;
 		wqe->recv.sgl[i].stag = 0;
 		wqe->recv.sgl[i].len = 0;
 		wqe->recv.sgl[i].to = 0;
+		wqe->recv.pbl_addr[i] = 0;
 	}
+	qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
+			     qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
+	qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
+			     qhp->wq.rq_size_log2)].pbl_addr = pbl_addr;
 	return 0;
 }
 
@@ -312,18 +396,18 @@ int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 			if (wr->send_flags & IB_SEND_FENCE)
 				t3_wr_flags |= T3_READ_FENCE_FLAG;
 			t3_wr_opcode = T3_WR_SEND;
-			err = iwch_build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
+			err = build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
 			break;
 		case IB_WR_RDMA_WRITE:
 		case IB_WR_RDMA_WRITE_WITH_IMM:
 			t3_wr_opcode = T3_WR_WRITE;
-			err = iwch_build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
+			err = build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
 			break;
 		case IB_WR_RDMA_READ:
 		case IB_WR_RDMA_READ_WITH_INV:
 			t3_wr_opcode = T3_WR_READ;
 			t3_wr_flags = 0; /* T3 reads are always signaled */
-			err = iwch_build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
+			err = build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
 			if (err)
 				break;
 			sqp->read_len = wqe->read.local_len;
@@ -332,14 +416,14 @@ int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 			break;
 		case IB_WR_FAST_REG_MR:
 			t3_wr_opcode = T3_WR_FASTREG;
-			err = iwch_build_fastreg(wqe, wr, &t3_wr_flit_cnt,
+			err = build_fastreg(wqe, wr, &t3_wr_flit_cnt,
 						 &wr_cnt, &qhp->wq);
 			break;
 		case IB_WR_LOCAL_INV:
 			if (wr->send_flags & IB_SEND_FENCE)
 				t3_wr_flags |= T3_LOCAL_FENCE_FLAG;
 			t3_wr_opcode = T3_WR_INV_STAG;
-			err = iwch_build_inv_stag(wqe, wr, &t3_wr_flit_cnt);
+			err = build_inv_stag(wqe, wr, &t3_wr_flit_cnt);
 			break;
 		default:
 			PDBG("%s post of type=%d TBD!\n", __func__,
@@ -398,18 +482,24 @@ int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 		return -EINVAL;
 	}
 	while (wr) {
+		if (wr->num_sge > T3_MAX_SGE) {
+			err = -EINVAL;
+			*bad_wr = wr;
+			break;
+		}
 		idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 		wqe = (union t3_wr *) (qhp->wq.queue + idx);
 		if (num_wrs)
-			err = iwch_build_rdma_recv(qhp->rhp, wqe, wr);
+			if (wr->sg_list[0].lkey)
+				err = build_rdma_recv(qhp, wqe, wr);
+			else
+				err = build_zero_stag_recv(qhp, wqe, wr);
 		else
 			err = -ENOMEM;
 		if (err) {
 			*bad_wr = wr;
 			break;
 		}
-		qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr, qhp->wq.rq_size_log2)] =
-			wr->wr_id;
 		build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
 			       Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
 			       0, sizeof(struct t3_receive_wr) >> 3, T3_SOPEOP);
@@ -810,7 +900,8 @@ static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
 	init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
 	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 (!qhp->ibqp.uobject)
+		init_attr.flags |= PRIV_QP;
 	if (peer2peer) {
 		init_attr.rtr_type = RTR_READ;
 		if (init_attr.ord == 0 && qhp->attr.mpa_attr.initiator)