RDMA/hns: Support direct wqe of userspace

The current write wqe mechanism is to write to DDR first, and then notify
the hardware through doorbell to read the data. Direct wqe is a mechanism
to fill wqe directly into the hardware. In the case of light load, the wqe
will be filled into pcie bar space of the hardware, this will reduce one
memory access operation and therefore reduce the latency. SIMD
instructions allows cpu to write the 512 bits at one time to device
memory, thus it can be used for posting direct wqe.

Add direct wqe enable switch and address mapping.

Link: https://lore.kernel.org/r/20211207124901.42123-2-liangwenpeng@huawei.com
Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

1 files changed, 53 insertions, 1 deletions

diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c
index 4fcab1611548..c84e1c23722c 100644
--- a/drivers/infiniband/hw/hns/hns_roce_qp.c
+++ b/drivers/infiniband/hw/hns/hns_roce_qp.c
@@ -379,6 +379,11 @@ err_out:
 	return ret;
 }
 
+static void qp_user_mmap_entry_remove(struct hns_roce_qp *hr_qp)
+{
+	rdma_user_mmap_entry_remove(&hr_qp->dwqe_mmap_entry->rdma_entry);
+}
+
 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
 {
 	struct xarray *xa = &hr_dev->qp_table_xa;
@@ -780,7 +785,11 @@ static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
 		goto err_inline;
 	}
 
+	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_DIRECT_WQE)
+		hr_qp->en_flags |= HNS_ROCE_QP_CAP_DIRECT_WQE;
+
 	return 0;
+
 err_inline:
 	free_rq_inline_buf(hr_qp);
 
@@ -822,6 +831,35 @@ static inline bool kernel_qp_has_rdb(struct hns_roce_dev *hr_dev,
 		hns_roce_qp_has_rq(init_attr));
 }
 
+static int qp_mmap_entry(struct hns_roce_qp *hr_qp,
+			 struct hns_roce_dev *hr_dev,
+			 struct ib_udata *udata,
+			 struct hns_roce_ib_create_qp_resp *resp)
+{
+	struct hns_roce_ucontext *uctx =
+		rdma_udata_to_drv_context(udata,
+			struct hns_roce_ucontext, ibucontext);
+	struct rdma_user_mmap_entry *rdma_entry;
+	u64 address;
+
+	address = hr_dev->dwqe_page + hr_qp->qpn * HNS_ROCE_DWQE_SIZE;
+
+	hr_qp->dwqe_mmap_entry =
+		hns_roce_user_mmap_entry_insert(&uctx->ibucontext, address,
+						HNS_ROCE_DWQE_SIZE,
+						HNS_ROCE_MMAP_TYPE_DWQE);
+
+	if (!hr_qp->dwqe_mmap_entry) {
+		ibdev_err(&hr_dev->ib_dev, "failed to get dwqe mmap entry.\n");
+		return -ENOMEM;
+	}
+
+	rdma_entry = &hr_qp->dwqe_mmap_entry->rdma_entry;
+	resp->dwqe_mmap_key = rdma_user_mmap_get_offset(rdma_entry);
+
+	return 0;
+}
+
 static int alloc_user_qp_db(struct hns_roce_dev *hr_dev,
 			    struct hns_roce_qp *hr_qp,
 			    struct ib_qp_init_attr *init_attr,
@@ -909,10 +947,16 @@ static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
 		hr_qp->en_flags |= HNS_ROCE_QP_CAP_OWNER_DB;
 
 	if (udata) {
+		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE) {
+			ret = qp_mmap_entry(hr_qp, hr_dev, udata, resp);
+			if (ret)
+				return ret;
+		}
+
 		ret = alloc_user_qp_db(hr_dev, hr_qp, init_attr, udata, ucmd,
 				       resp);
 		if (ret)
-			return ret;
+			goto err_remove_qp;
 	} else {
 		ret = alloc_kernel_qp_db(hr_dev, hr_qp, init_attr);
 		if (ret)
@@ -920,6 +964,12 @@ static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
 	}
 
 	return 0;
+
+err_remove_qp:
+	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
+		qp_user_mmap_entry_remove(hr_qp);
+
+	return ret;
 }
 
 static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
@@ -933,6 +983,8 @@ static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
 			hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB)
 			hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
+		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE)
+			qp_user_mmap_entry_remove(hr_qp);
 	} else {
 		if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
 			hns_roce_free_db(hr_dev, &hr_qp->rdb);