RDMA/hns: Support adaptive hopnum for MTR

driver inclusion
category: feature
bugzilla: https://gitee.com/openeuler/kernel/issues/I5Y79T

-------------------------------------------------------------------

In the current implementation, we configure a fixed addressing level
for the MTR. But in fact, the necessary addressing level is related
to the page size and the size of the memory.

This patch calculates the addressing level according to the page size
and the size of the memory, and uses the addressing level to configure
the MTR.

It only works for PBL now.
Signed-off-by: NChengchang Tang <tangchengchang@huawei.com>
Reviewed-by: NYangyang Li <liyangyang20@huawei.com>
Reviewed-by: NYueHaibing <yuehaibing@huawei.com>
Signed-off-by: NZheng Zengkai <zhengzengkai@huawei.com>

drivers/infiniband/hw/hns/hns_roce_device.h

@@ -178,6 +178,8 @@ enum {
 
 #define HNS_ROCE_CMD_SUCCESS			1
 
+#define HNS_ROCE_MAX_HOP_NUM			3
+
 /* The minimum page size is 4K for hardware */
 #define HNS_HW_PAGE_SHIFT			12
 #define HNS_HW_PAGE_SIZE			(1 << HNS_HW_PAGE_SHIFT)

drivers/infiniband/hw/hns/hns_roce_mr.c

@@ -110,10 +110,13 @@ static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr,
 	err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr,
 				  hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT,
 				  udata, start);
-	if (err)
+	if (err) {
 		ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err);
-	else
-		mr->npages = mr->pbl_mtr.hem_cfg.buf_pg_count;
+		return err;
+	}
+
+	mr->npages = mr->pbl_mtr.hem_cfg.buf_pg_count;
+	mr->pbl_hop_num = buf_attr.region[0].hopnum;
 
 	return err;
 }
@@ -943,6 +946,52 @@ static int get_best_page_shift(struct hns_roce_dev *hr_dev,
 	return 0;
 }
 
+static int get_best_hop_num(struct hns_roce_dev *hr_dev,
+			    struct hns_roce_mtr *mtr,
+			    struct hns_roce_buf_attr *buf_attr,
+			    unsigned int ba_pg_shift)
+{
+#define INVALID_HOPNUM -1
+#define MIN_BA_CNT 1
+	size_t buf_pg_sz = 1 << buf_attr->page_shift;
+	struct ib_device *ibdev = &hr_dev->ib_dev;
+	size_t ba_pg_sz = 1 << ba_pg_shift;
+	int hop_num = INVALID_HOPNUM;
+	size_t unit = MIN_BA_CNT;
+	size_t ba_cnt;
+	int i;
+
+	if (!buf_attr->adaptive)
+		return 0;
+
+	hop_num = INVALID_HOPNUM;
+	unit = MIN_BA_CNT;
+	/* Caculating the number of buf pages, each buf page needs a BA */
+	if (mtr->umem)
+		ba_cnt = ib_umem_num_dma_blocks(mtr->umem, buf_pg_sz);
+	else
+		ba_cnt = DIV_ROUND_UP(buf_attr->region[0].size, buf_pg_sz);
+
+	for (i = 0; i <= HNS_ROCE_MAX_HOP_NUM; i++) {
+		if (ba_cnt <= unit) {
+			hop_num = i;
+			break;
+		}
+		/* Number of BAs can be represented at per hop */
+		unit *= ba_pg_sz / BA_BYTE_LEN;
+	}
+
+	if (hop_num < 0) {
+		ibdev_err(ibdev,
+			  "failed to calculate a valid hopnum.\n");
+		return -EINVAL;
+	}
+
+	buf_attr->region[0].hopnum = hop_num;
+
+	return 0;
+}
+
 static bool is_buf_attr_valid(struct hns_roce_dev *hr_dev,
 			      struct hns_roce_buf_attr *attr)
 {
@@ -1068,6 +1117,10 @@ int hns_roce_mtr_create(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
 		ret = get_best_page_shift(hr_dev, mtr, buf_attr);
 		if (ret)
 			goto err_init_buf;
+
+		ret = get_best_hop_num(hr_dev, mtr, buf_attr, ba_page_shift);
+		if (ret)
+			goto err_init_buf;
 	}
 
 	ret = mtr_init_buf_cfg(hr_dev, mtr, buf_attr);