[PATCH for-next 0/2] RDMA/hns: Optimize codes related to multi-hop addressing

[PATCH for-next 0/2] RDMA/hns: Optimize codes related to multi-hop addressing

[Weihang Li]

Codes about getting/putting HEM(Hardware Entry Memory) in multi-hop
addressing are not clear enough, so optimize by spliting and encapsulating
them into new functions.

Xi Wang (2):
  RDMA/hns: Optimize mhop get flow for multi-hop addressing
  RDMA/hns: Optimize mhop put flow for multi-hop addressing

 drivers/infiniband/hw/hns/hns_roce_hem.c | 458 ++++++++++++++++---------------
 1 file changed, 243 insertions(+), 215 deletions(-)

-- 
2.8.1

[PATCH for-next 1/2] RDMA/hns: Optimize mhop get flow for multi-hop addressing

[Weihang Li]

From: Xi Wang <wangxi11@huawei.com>

Splits hns_roce_table_mhop_get() into 4 sub-functions to make the code flow
clearer.

Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
 drivers/infiniband/hw/hns/hns_roce_hem.c | 297 +++++++++++++++++++------------
 1 file changed, 182 insertions(+), 115 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.c b/drivers/infiniband/hw/hns/hns_roce_hem.c
index 8380d71..cc557f1 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hem.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hem.c
@@ -39,6 +39,16 @@
 #define DMA_ADDR_T_SHIFT		12
 #define BT_BA_SHIFT			32
 
+#define HEM_INDEX_BUF			BIT(0)
+#define HEM_INDEX_L0			BIT(1)
+#define HEM_INDEX_L1			BIT(2)
+struct hns_roce_hem_index {
+	u64 buf;
+	u64 l0;
+	u64 l1;
+	u32 inited; /* indicate which index is available */
+};
+
 bool hns_roce_check_whether_mhop(struct hns_roce_dev *hr_dev, u32 type)
 {
 	int hop_num = 0;
@@ -434,178 +444,235 @@ static int hns_roce_set_hem(struct hns_roce_dev *hr_dev,
 	return ret;
 }
 
-static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,
-				   struct hns_roce_hem_table *table,
-				   unsigned long obj)
+static int calc_hem_config(struct hns_roce_dev *hr_dev,
+			   struct hns_roce_hem_table *table, unsigned long obj,
+			   struct hns_roce_hem_mhop *mhop,
+			   struct hns_roce_hem_index *index)
 {
-	struct device *dev = hr_dev->dev;
-	struct hns_roce_hem_mhop mhop;
-	struct hns_roce_hem_iter iter;
-	u32 buf_chunk_size;
-	u32 bt_chunk_size;
+	struct ib_device *ibdev = &hr_dev->ib_dev;
+	unsigned long mhop_obj = obj;
+	u32 l0_idx, l1_idx, l2_idx;
 	u32 chunk_ba_num;
-	u32 hop_num;
-	u32 size;
 	u32 bt_num;
-	u64 hem_idx;
-	u64 bt_l1_idx = 0;
-	u64 bt_l0_idx = 0;
-	u64 bt_ba;
-	unsigned long mhop_obj = obj;
-	int bt_l1_allocated = 0;
-	int bt_l0_allocated = 0;
-	int step_idx;
 	int ret;
 
-	ret = hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
+	ret = hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, mhop);
 	if (ret)
 		return ret;
 
-	buf_chunk_size = mhop.buf_chunk_size;
-	bt_chunk_size = mhop.bt_chunk_size;
-	hop_num = mhop.hop_num;
-	chunk_ba_num = bt_chunk_size / BA_BYTE_LEN;
-
-	bt_num = hns_roce_get_bt_num(table->type, hop_num);
+	l0_idx = mhop->l0_idx;
+	l1_idx = mhop->l1_idx;
+	l2_idx = mhop->l2_idx;
+	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
+	bt_num = hns_roce_get_bt_num(table->type, mhop->hop_num);
 	switch (bt_num) {
 	case 3:
-		hem_idx = mhop.l0_idx * chunk_ba_num * chunk_ba_num +
-			  mhop.l1_idx * chunk_ba_num + mhop.l2_idx;
-		bt_l1_idx = mhop.l0_idx * chunk_ba_num + mhop.l1_idx;
-		bt_l0_idx = mhop.l0_idx;
+		index->l1 = l0_idx * chunk_ba_num + l1_idx;
+		index->l0 = l0_idx;
+		index->buf = l0_idx * chunk_ba_num * chunk_ba_num +
+			     l1_idx * chunk_ba_num + l2_idx;
 		break;
 	case 2:
-		hem_idx = mhop.l0_idx * chunk_ba_num + mhop.l1_idx;
-		bt_l0_idx = mhop.l0_idx;
+		index->l0 = l0_idx;
+		index->buf = l0_idx * chunk_ba_num + l1_idx;
 		break;
 	case 1:
-		hem_idx = mhop.l0_idx;
+		index->buf = l0_idx;
 		break;
 	default:
-		dev_err(dev, "Table %d not support hop_num = %d!\n",
-			     table->type, hop_num);
+		ibdev_err(ibdev, "Table %d not support mhop.hop_num = %d!\n",
+			  table->type, mhop->hop_num);
 		return -EINVAL;
 	}
 
-	if (unlikely(hem_idx >= table->num_hem)) {
-		dev_err(dev, "Table %d exceed hem limt idx = %llu,max = %lu!\n",
-			     table->type, hem_idx, table->num_hem);
+	if (unlikely(index->buf >= table->num_hem)) {
+		ibdev_err(ibdev, "Table %d exceed hem limt idx %llu,max %lu!\n",
+			  table->type, index->buf, table->num_hem);
 		return -EINVAL;
 	}
 
-	mutex_lock(&table->mutex);
+	return 0;
+}
 
-	if (table->hem[hem_idx]) {
-		++table->hem[hem_idx]->refcount;
-		goto out;
+static void free_mhop_hem(struct hns_roce_dev *hr_dev,
+			  struct hns_roce_hem_table *table,
+			  struct hns_roce_hem_mhop *mhop,
+			  struct hns_roce_hem_index *index)
+{
+	u32 bt_size = mhop->bt_chunk_size;
+	struct device *dev = hr_dev->dev;
+
+	if (index->inited & HEM_INDEX_BUF) {
+		hns_roce_free_hem(hr_dev, table->hem[index->buf]);
+		table->hem[index->buf] = NULL;
+	}
+
+	if (index->inited & HEM_INDEX_L1) {
+		dma_free_coherent(dev, bt_size, table->bt_l1[index->l1],
+				  table->bt_l1_dma_addr[index->l1]);
+		table->bt_l1[index->l1] = NULL;
 	}
 
+	if (index->inited & HEM_INDEX_L0) {
+		dma_free_coherent(dev, bt_size, table->bt_l0[index->l0],
+				  table->bt_l0_dma_addr[index->l0]);
+		table->bt_l0[index->l0] = NULL;
+	}
+}
+
+static int alloc_mhop_hem(struct hns_roce_dev *hr_dev,
+			  struct hns_roce_hem_table *table,
+			  struct hns_roce_hem_mhop *mhop,
+			  struct hns_roce_hem_index *index)
+{
+	u32 bt_size = mhop->bt_chunk_size;
+	struct device *dev = hr_dev->dev;
+	struct hns_roce_hem_iter iter;
+	gfp_t flag;
+	u64 bt_ba;
+	u32 size;
+	int ret;
+
 	/* alloc L1 BA's chunk */
-	if ((check_whether_bt_num_3(table->type, hop_num) ||
-		check_whether_bt_num_2(table->type, hop_num)) &&
-		!table->bt_l0[bt_l0_idx]) {
-		table->bt_l0[bt_l0_idx] = dma_alloc_coherent(dev, bt_chunk_size,
-					    &(table->bt_l0_dma_addr[bt_l0_idx]),
+	if ((check_whether_bt_num_3(table->type, mhop->hop_num) ||
+	     check_whether_bt_num_2(table->type, mhop->hop_num)) &&
+	     !table->bt_l0[index->l0]) {
+		table->bt_l0[index->l0] = dma_alloc_coherent(dev, bt_size,
+					    &table->bt_l0_dma_addr[index->l0],
 					    GFP_KERNEL);
-		if (!table->bt_l0[bt_l0_idx]) {
+		if (!table->bt_l0[index->l0]) {
 			ret = -ENOMEM;
 			goto out;
 		}
-		bt_l0_allocated = 1;
-
-		/* set base address to hardware */
-		if (table->type < HEM_TYPE_MTT) {
-			step_idx = 0;
-			if (hr_dev->hw->set_hem(hr_dev, table, obj, step_idx)) {
-				ret = -ENODEV;
-				dev_err(dev, "set HEM base address to HW failed!\n");
-				goto err_dma_alloc_l1;
-			}
-		}
+		index->inited |= HEM_INDEX_L0;
 	}
 
 	/* alloc L2 BA's chunk */
-	if (check_whether_bt_num_3(table->type, hop_num) &&
-	    !table->bt_l1[bt_l1_idx])  {
-		table->bt_l1[bt_l1_idx] = dma_alloc_coherent(dev, bt_chunk_size,
-					    &(table->bt_l1_dma_addr[bt_l1_idx]),
+	if (check_whether_bt_num_3(table->type, mhop->hop_num) &&
+	    !table->bt_l1[index->l1])  {
+		table->bt_l1[index->l1] = dma_alloc_coherent(dev, bt_size,
+					    &table->bt_l1_dma_addr[index->l1],
 					    GFP_KERNEL);
-		if (!table->bt_l1[bt_l1_idx]) {
+		if (!table->bt_l1[index->l1]) {
 			ret = -ENOMEM;
-			goto err_dma_alloc_l1;
-		}
-		bt_l1_allocated = 1;
-		*(table->bt_l0[bt_l0_idx] + mhop.l1_idx) =
-					       table->bt_l1_dma_addr[bt_l1_idx];
-
-		/* set base address to hardware */
-		step_idx = 1;
-		if (hr_dev->hw->set_hem(hr_dev, table, obj, step_idx)) {
-			ret = -ENODEV;
-			dev_err(dev, "set HEM base address to HW failed!\n");
-			goto err_alloc_hem_buf;
+			goto err_alloc_hem;
 		}
+		index->inited |= HEM_INDEX_L1;
+		*(table->bt_l0[index->l0] + mhop->l1_idx) =
+					       table->bt_l1_dma_addr[index->l1];
 	}
 
 	/*
 	 * alloc buffer space chunk for QPC/MTPT/CQC/SRQC/SCCC.
 	 * alloc bt space chunk for MTT/CQE.
 	 */
-	size = table->type < HEM_TYPE_MTT ? buf_chunk_size : bt_chunk_size;
-	table->hem[hem_idx] = hns_roce_alloc_hem(hr_dev,
-						size >> PAGE_SHIFT,
-						size,
-						(table->lowmem ? GFP_KERNEL :
-						GFP_HIGHUSER) | __GFP_NOWARN);
-	if (!table->hem[hem_idx]) {
+	size = table->type < HEM_TYPE_MTT ? mhop->buf_chunk_size : bt_size;
+	flag = (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) | __GFP_NOWARN;
+	table->hem[index->buf] = hns_roce_alloc_hem(hr_dev, size >> PAGE_SHIFT,
+						    size, flag);
+	if (!table->hem[index->buf]) {
 		ret = -ENOMEM;
-		goto err_alloc_hem_buf;
+		goto err_alloc_hem;
 	}
 
-	hns_roce_hem_first(table->hem[hem_idx], &iter);
+	index->inited |= HEM_INDEX_BUF;
+	hns_roce_hem_first(table->hem[index->buf], &iter);
 	bt_ba = hns_roce_hem_addr(&iter);
-
 	if (table->type < HEM_TYPE_MTT) {
-		if (hop_num == 2) {
-			*(table->bt_l1[bt_l1_idx] + mhop.l2_idx) = bt_ba;
-			step_idx = 2;
-		} else if (hop_num == 1) {
-			*(table->bt_l0[bt_l0_idx] + mhop.l1_idx) = bt_ba;
-			step_idx = 1;
-		} else if (hop_num == HNS_ROCE_HOP_NUM_0) {
-			step_idx = 0;
-		} else {
-			ret = -EINVAL;
-			goto err_dma_alloc_l1;
+		if (mhop->hop_num == 2)
+			*(table->bt_l1[index->l1] + mhop->l2_idx) = bt_ba;
+		else if (mhop->hop_num == 1)
+			*(table->bt_l0[index->l0] + mhop->l1_idx) = bt_ba;
+	} else if (mhop->hop_num == 2) {
+		*(table->bt_l0[index->l0] + mhop->l1_idx) = bt_ba;
+	}
+
+	return 0;
+err_alloc_hem:
+	free_mhop_hem(hr_dev, table, mhop, index);
+out:
+	return ret;
+}
+
+static int set_mhop_hem(struct hns_roce_dev *hr_dev,
+			struct hns_roce_hem_table *table, unsigned long obj,
+			struct hns_roce_hem_mhop *mhop,
+			struct hns_roce_hem_index *index)
+{
+	struct ib_device *ibdev = &hr_dev->ib_dev;
+	int step_idx;
+	int ret;
+
+	if (index->inited & HEM_INDEX_L0) {
+		ret = hr_dev->hw->set_hem(hr_dev, table, obj, 0);
+		if (ret) {
+			ibdev_err(ibdev, "set HEM step 0 failed!\n");
+			goto out;
 		}
+	}
 
-		/* set HEM base address to hardware */
-		if (hr_dev->hw->set_hem(hr_dev, table, obj, step_idx)) {
-			ret = -ENODEV;
-			dev_err(dev, "set HEM base address to HW failed!\n");
-			goto err_alloc_hem_buf;
+	if (index->inited & HEM_INDEX_L1) {
+		ret = hr_dev->hw->set_hem(hr_dev, table, obj, 1);
+		if (ret) {
+			ibdev_err(ibdev, "set HEM step 1 failed!\n");
+			goto out;
 		}
-	} else if (hop_num == 2) {
-		*(table->bt_l0[bt_l0_idx] + mhop.l1_idx) = bt_ba;
 	}
 
-	++table->hem[hem_idx]->refcount;
-	goto out;
+	if (index->inited & HEM_INDEX_BUF) {
+		if (mhop->hop_num == HNS_ROCE_HOP_NUM_0)
+			step_idx = 0;
+		else
+			step_idx = mhop->hop_num;
+		ret = hr_dev->hw->set_hem(hr_dev, table, obj, step_idx);
+		if (ret)
+			ibdev_err(ibdev, "set HEM step last failed!\n");
+	}
+out:
+	return ret;
+}
 
-err_alloc_hem_buf:
-	if (bt_l1_allocated) {
-		dma_free_coherent(dev, bt_chunk_size, table->bt_l1[bt_l1_idx],
-				  table->bt_l1_dma_addr[bt_l1_idx]);
-		table->bt_l1[bt_l1_idx] = NULL;
+static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,
+				   struct hns_roce_hem_table *table,
+				   unsigned long obj)
+{
+	struct ib_device *ibdev = &hr_dev->ib_dev;
+	struct hns_roce_hem_index index = {};
+	struct hns_roce_hem_mhop mhop = {};
+	int ret;
+
+	ret = calc_hem_config(hr_dev, table, obj, &mhop, &index);
+	if (ret) {
+		ibdev_err(ibdev, "calc hem config failed!\n");
+		return ret;
 	}
 
-err_dma_alloc_l1:
-	if (bt_l0_allocated) {
-		dma_free_coherent(dev, bt_chunk_size, table->bt_l0[bt_l0_idx],
-				  table->bt_l0_dma_addr[bt_l0_idx]);
-		table->bt_l0[bt_l0_idx] = NULL;
+	mutex_lock(&table->mutex);
+	if (table->hem[index.buf]) {
+		++table->hem[index.buf]->refcount;
+		goto out;
+	}
+
+	ret = alloc_mhop_hem(hr_dev, table, &mhop, &index);
+	if (ret) {
+		ibdev_err(ibdev, "alloc mhop hem failed!\n");
+		goto out;
+	}
+
+	/* set HEM base address to hardware */
+	if (table->type < HEM_TYPE_MTT) {
+		ret = set_mhop_hem(hr_dev, table, obj, &mhop, &index);
+		if (ret) {
+			ibdev_err(ibdev, "set HEM address to HW failed!\n");
+			goto err_alloc;
+		}
 	}
 
+	++table->hem[index.buf]->refcount;
+	goto out;
+
+err_alloc:
+	free_mhop_hem(hr_dev, table, &mhop, &index);
 out:
 	mutex_unlock(&table->mutex);
 	return ret;
-- 
2.8.1

[PATCH for-next 2/2] RDMA/hns: Optimize mhop put flow for multi-hop addressing

[Weihang Li]

From: Xi Wang <wangxi11@huawei.com>

Optimizes hns_roce_table_mhop_get() by encapsulating code about clearing
hem into clear_mhop_hem(), which will make the code flow clearer.

Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
 drivers/infiniband/hw/hns/hns_roce_hem.c | 161 ++++++++++++-------------------
 1 file changed, 61 insertions(+), 100 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.c b/drivers/infiniband/hw/hns/hns_roce_hem.c
index cc557f1..c963787 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hem.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hem.c
@@ -94,25 +94,27 @@ bool hns_roce_check_whether_mhop(struct hns_roce_dev *hr_dev, u32 type)
 	return hop_num ? true : false;
 }
 
-static bool hns_roce_check_hem_null(struct hns_roce_hem **hem, u64 start_idx,
-			    u32 bt_chunk_num, u64 hem_max_num)
+static bool hns_roce_check_hem_null(struct hns_roce_hem **hem, u64 hem_idx,
+				    u32 bt_chunk_num, u64 hem_max_num)
 {
+	u64 start_idx = round_down(hem_idx, bt_chunk_num);
 	u64 check_max_num = start_idx + bt_chunk_num;
 	u64 i;
 
 	for (i = start_idx; (i < check_max_num) && (i < hem_max_num); i++)
-		if (hem[i])
+		if (i != hem_idx && hem[i])
 			return false;
 
 	return true;
 }
 
-static bool hns_roce_check_bt_null(u64 **bt, u64 start_idx, u32 bt_chunk_num)
+static bool hns_roce_check_bt_null(u64 **bt, u64 ba_idx, u32 bt_chunk_num)
 {
+	u64 start_idx = round_down(ba_idx, bt_chunk_num);
 	int i;
 
 	for (i = 0; i < bt_chunk_num; i++)
-		if (bt[start_idx + i])
+		if (i != ba_idx && bt[start_idx + i])
 			return false;
 
 	return true;
@@ -723,116 +725,75 @@ int hns_roce_table_get(struct hns_roce_dev *hr_dev,
 	return ret;
 }
 
+static void clear_mhop_hem(struct hns_roce_dev *hr_dev,
+			   struct hns_roce_hem_table *table, unsigned long obj,
+			   struct hns_roce_hem_mhop *mhop,
+			   struct hns_roce_hem_index *index)
+{
+	struct ib_device *ibdev = &hr_dev->ib_dev;
+	u32 hop_num = mhop->hop_num;
+	u32 chunk_ba_num;
+	int step_idx;
+
+	index->inited = HEM_INDEX_BUF;
+	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
+	if (check_whether_bt_num_2(table->type, hop_num)) {
+		if (hns_roce_check_hem_null(table->hem, index->buf,
+					    chunk_ba_num, table->num_hem))
+			index->inited |= HEM_INDEX_L0;
+	} else if (check_whether_bt_num_3(table->type, hop_num)) {
+		if (hns_roce_check_hem_null(table->hem, index->buf,
+					    chunk_ba_num, table->num_hem)) {
+			index->inited |= HEM_INDEX_L1;
+			if (hns_roce_check_bt_null(table->bt_l1, index->l1,
+						   chunk_ba_num))
+				index->inited |= HEM_INDEX_L0;
+		}
+	}
+
+	if (table->type < HEM_TYPE_MTT) {
+		if (hop_num == HNS_ROCE_HOP_NUM_0)
+			step_idx = 0;
+		else
+			step_idx = hop_num;
+
+		if (hr_dev->hw->clear_hem(hr_dev, table, obj, step_idx))
+			ibdev_warn(ibdev, "Clear hop%d HEM failed.\n", hop_num);
+
+		if (index->inited & HEM_INDEX_L1)
+			if (hr_dev->hw->clear_hem(hr_dev, table, obj, 1))
+				ibdev_warn(ibdev, "Clear HEM step 1 failed.\n");
+
+		if (index->inited & HEM_INDEX_L0)
+			if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
+				ibdev_warn(ibdev, "Clear HEM step 0 failed.\n");
+	}
+}
+
 static void hns_roce_table_mhop_put(struct hns_roce_dev *hr_dev,
 				    struct hns_roce_hem_table *table,
 				    unsigned long obj,
 				    int check_refcount)
 {
-	struct device *dev = hr_dev->dev;
-	struct hns_roce_hem_mhop mhop;
-	unsigned long mhop_obj = obj;
-	u32 bt_chunk_size;
-	u32 chunk_ba_num;
-	u32 hop_num;
-	u32 start_idx;
-	u32 bt_num;
-	u64 hem_idx;
-	u64 bt_l1_idx = 0;
+	struct ib_device *ibdev = &hr_dev->ib_dev;
+	struct hns_roce_hem_index index = {};
+	struct hns_roce_hem_mhop mhop = {};
 	int ret;
 
-	ret = hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
-	if (ret)
-		return;
-
-	bt_chunk_size = mhop.bt_chunk_size;
-	hop_num = mhop.hop_num;
-	chunk_ba_num = bt_chunk_size / BA_BYTE_LEN;
-
-	bt_num = hns_roce_get_bt_num(table->type, hop_num);
-	switch (bt_num) {
-	case 3:
-		hem_idx = mhop.l0_idx * chunk_ba_num * chunk_ba_num +
-			  mhop.l1_idx * chunk_ba_num + mhop.l2_idx;
-		bt_l1_idx = mhop.l0_idx * chunk_ba_num + mhop.l1_idx;
-		break;
-	case 2:
-		hem_idx = mhop.l0_idx * chunk_ba_num + mhop.l1_idx;
-		break;
-	case 1:
-		hem_idx = mhop.l0_idx;
-		break;
-	default:
-		dev_err(dev, "Table %d not support hop_num = %d!\n",
-			     table->type, hop_num);
+	ret = calc_hem_config(hr_dev, table, obj, &mhop, &index);
+	if (ret) {
+		ibdev_err(ibdev, "calc hem config failed!\n");
 		return;
 	}
 
 	mutex_lock(&table->mutex);
-
-	if (check_refcount && (--table->hem[hem_idx]->refcount > 0)) {
+	if (check_refcount && (--table->hem[index.buf]->refcount > 0)) {
 		mutex_unlock(&table->mutex);
 		return;
 	}
 
-	if (table->type < HEM_TYPE_MTT && hop_num == 1) {
-		if (hr_dev->hw->clear_hem(hr_dev, table, obj, 1))
-			dev_warn(dev, "Clear HEM base address failed.\n");
-	} else if (table->type < HEM_TYPE_MTT && hop_num == 2) {
-		if (hr_dev->hw->clear_hem(hr_dev, table, obj, 2))
-			dev_warn(dev, "Clear HEM base address failed.\n");
-	} else if (table->type < HEM_TYPE_MTT &&
-		   hop_num == HNS_ROCE_HOP_NUM_0) {
-		if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
-			dev_warn(dev, "Clear HEM base address failed.\n");
-	}
-
-	/*
-	 * free buffer space chunk for QPC/MTPT/CQC/SRQC/SCCC.
-	 * free bt space chunk for MTT/CQE.
-	 */
-	hns_roce_free_hem(hr_dev, table->hem[hem_idx]);
-	table->hem[hem_idx] = NULL;
-
-	if (check_whether_bt_num_2(table->type, hop_num)) {
-		start_idx = mhop.l0_idx * chunk_ba_num;
-		if (hns_roce_check_hem_null(table->hem, start_idx,
-					    chunk_ba_num, table->num_hem)) {
-			if (table->type < HEM_TYPE_MTT &&
-			    hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
-				dev_warn(dev, "Clear HEM base address failed.\n");
-
-			dma_free_coherent(dev, bt_chunk_size,
-					  table->bt_l0[mhop.l0_idx],
-					  table->bt_l0_dma_addr[mhop.l0_idx]);
-			table->bt_l0[mhop.l0_idx] = NULL;
-		}
-	} else if (check_whether_bt_num_3(table->type, hop_num)) {
-		start_idx = mhop.l0_idx * chunk_ba_num * chunk_ba_num +
-			    mhop.l1_idx * chunk_ba_num;
-		if (hns_roce_check_hem_null(table->hem, start_idx,
-					    chunk_ba_num, table->num_hem)) {
-			if (hr_dev->hw->clear_hem(hr_dev, table, obj, 1))
-				dev_warn(dev, "Clear HEM base address failed.\n");
-
-			dma_free_coherent(dev, bt_chunk_size,
-					  table->bt_l1[bt_l1_idx],
-					  table->bt_l1_dma_addr[bt_l1_idx]);
-			table->bt_l1[bt_l1_idx] = NULL;
-
-			start_idx = mhop.l0_idx * chunk_ba_num;
-			if (hns_roce_check_bt_null(table->bt_l1, start_idx,
-						   chunk_ba_num)) {
-				if (hr_dev->hw->clear_hem(hr_dev, table, obj,
-							  0))
-					dev_warn(dev, "Clear HEM base address failed.\n");
-
-				dma_free_coherent(dev, bt_chunk_size,
-					    table->bt_l0[mhop.l0_idx],
-					    table->bt_l0_dma_addr[mhop.l0_idx]);
-				table->bt_l0[mhop.l0_idx] = NULL;
-			}
-		}
-	}
+	clear_mhop_hem(hr_dev, table, obj, &mhop, &index);
+	free_mhop_hem(hr_dev, table, &mhop, &index);
 
 	mutex_unlock(&table->mutex);
 }
-- 
2.8.1

Re: [PATCH for-next 0/2] RDMA/hns: Optimize codes related to multi-hop addressing

[Jason Gunthorpe]

On Tue, Mar 17, 2020 at 11:55:22AM +0800, Weihang Li wrote:
> Codes about getting/putting HEM(Hardware Entry Memory) in multi-hop
> addressing are not clear enough, so optimize by spliting and encapsulating
> them into new functions.
> 
> Xi Wang (2):
>   RDMA/hns: Optimize mhop get flow for multi-hop addressing
>   RDMA/hns: Optimize mhop put flow for multi-hop addressing

Applied to for-next, thanks

Jason