RDMA/hns: Add timer allocation support for hip08

This patch adds qpc timer and cqc timer allocation support for hardware
timeout retransmission in kernel space driver.
Signed-off-by: Yangyang Li <liyangyang20@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

drivers/infiniband/hw/hns/hns_roce_cmd.h

@@ -75,6 +75,10 @@ enum {
 	HNS_ROCE_CMD_DESTROY_MPT_BT1	= 0x29,
 	HNS_ROCE_CMD_DESTROY_MPT_BT2	= 0x2a,
 
+	/* CQC TIMER commands */
+	HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0 = 0x23,
+	HNS_ROCE_CMD_READ_CQC_TIMER_BT0  = 0x27,
+
 	/* MPT commands */
 	HNS_ROCE_CMD_QUERY_MPT		= 0x62,
 
@@ -89,6 +93,10 @@ enum {
 	HNS_ROCE_CMD_DESTROY_SRQC_BT1	= 0x39,
 	HNS_ROCE_CMD_DESTROY_SRQC_BT2	= 0x3a,
 
+	/* QPC TIMER commands */
+	HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0 = 0x33,
+	HNS_ROCE_CMD_READ_QPC_TIMER_BT0  = 0x37,
+
 	/* EQC commands */
 	HNS_ROCE_CMD_CREATE_AEQC	= 0x80,
 	HNS_ROCE_CMD_MODIFY_AEQC	= 0x81,

drivers/infiniband/hw/hns/hns_roce_device.h

@@ -732,6 +732,8 @@ struct hns_roce_caps {
 	u32		max_extend_sg;
 	int		num_qps;	/* 256k */
 	int             reserved_qps;
+	int		num_qpc_timer;
+	int		num_cqc_timer;
 	u32		max_srq_sg;
 	int		num_srqs;
 	u32		max_wqes;	/* 16k */
@@ -772,6 +774,8 @@ struct hns_roce_caps {
 	int		trrl_entry_sz;
 	int		cqc_entry_sz;
 	int		sccc_entry_sz;
+	int		qpc_timer_entry_sz;
+	int		cqc_timer_entry_sz;
 	int		srqc_entry_sz;
 	int		idx_entry_sz;
 	u32		pbl_ba_pg_sz;
@@ -781,8 +785,10 @@ struct hns_roce_caps {
 	int		ceqe_depth;
 	enum ib_mtu	max_mtu;
 	u32		qpc_bt_num;
+	u32		qpc_timer_bt_num;
 	u32		srqc_bt_num;
 	u32		cqc_bt_num;
+	u32		cqc_timer_bt_num;
 	u32		mpt_bt_num;
 	u32		sccc_bt_num;
 	u32		qpc_ba_pg_sz;
@@ -803,6 +809,12 @@ struct hns_roce_caps {
 	u32		sccc_ba_pg_sz;
 	u32		sccc_buf_pg_sz;
 	u32		sccc_hop_num;
+	u32		qpc_timer_ba_pg_sz;
+	u32		qpc_timer_buf_pg_sz;
+	u32		qpc_timer_hop_num;
+	u32		cqc_timer_ba_pg_sz;
+	u32		cqc_timer_buf_pg_sz;
+	u32		cqc_timer_hop_num;
 	u32		cqe_ba_pg_sz;
 	u32		cqe_buf_pg_sz;
 	u32		cqe_hop_num;
@@ -932,6 +944,8 @@ struct hns_roce_dev {
 	struct hns_roce_srq_table srq_table;
 	struct hns_roce_qp_table  qp_table;
 	struct hns_roce_eq_table  eq_table;
+	struct hns_roce_hem_table  qpc_timer_table;
+	struct hns_roce_hem_table  cqc_timer_table;
 
 	int			cmd_mod;
 	int			loop_idc;

drivers/infiniband/hw/hns/hns_roce_hem.c

@@ -46,6 +46,8 @@ bool hns_roce_check_whether_mhop(struct hns_roce_dev *hr_dev, u32 type)
 	    (hr_dev->caps.cqc_hop_num && type == HEM_TYPE_CQC) ||
 	    (hr_dev->caps.srqc_hop_num && type == HEM_TYPE_SRQC) ||
 	    (hr_dev->caps.sccc_hop_num && type == HEM_TYPE_SCCC) ||
+	    (hr_dev->caps.qpc_timer_hop_num && type == HEM_TYPE_QPC_TIMER) ||
+	    (hr_dev->caps.cqc_timer_hop_num && type == HEM_TYPE_CQC_TIMER) ||
 	    (hr_dev->caps.cqe_hop_num && type == HEM_TYPE_CQE) ||
 	    (hr_dev->caps.mtt_hop_num && type == HEM_TYPE_MTT) ||
 	    (hr_dev->caps.srqwqe_hop_num && type == HEM_TYPE_SRQWQE) ||
@@ -134,6 +136,22 @@ int hns_roce_calc_hem_mhop(struct hns_roce_dev *hr_dev,
 		mhop->ba_l0_num = hr_dev->caps.sccc_bt_num;
 		mhop->hop_num = hr_dev->caps.sccc_hop_num;
 		break;
+	case HEM_TYPE_QPC_TIMER:
+		mhop->buf_chunk_size = 1 << (hr_dev->caps.qpc_timer_buf_pg_sz
+					     + PAGE_SHIFT);
+		mhop->bt_chunk_size = 1 << (hr_dev->caps.qpc_timer_ba_pg_sz
+					    + PAGE_SHIFT);
+		mhop->ba_l0_num = hr_dev->caps.qpc_timer_bt_num;
+		mhop->hop_num = hr_dev->caps.qpc_timer_hop_num;
+		break;
+	case HEM_TYPE_CQC_TIMER:
+		mhop->buf_chunk_size = 1 << (hr_dev->caps.cqc_timer_buf_pg_sz
+					     + PAGE_SHIFT);
+		mhop->bt_chunk_size = 1 << (hr_dev->caps.cqc_timer_ba_pg_sz
+					    + PAGE_SHIFT);
+		mhop->ba_l0_num = hr_dev->caps.cqc_timer_bt_num;
+		mhop->hop_num = hr_dev->caps.cqc_timer_hop_num;
+		break;
 	case HEM_TYPE_SRQC:
 		mhop->buf_chunk_size = 1 << (hr_dev->caps.srqc_buf_pg_sz
 					     + PAGE_SHIFT);
@@ -602,6 +620,7 @@ int hns_roce_table_get(struct hns_roce_dev *hr_dev,
 	mutex_unlock(&table->mutex);
 	return ret;
 }
+EXPORT_SYMBOL_GPL(hns_roce_table_get);
 
 static void hns_roce_table_mhop_put(struct hns_roce_dev *hr_dev,
 				    struct hns_roce_hem_table *table,
@@ -744,6 +763,7 @@ void hns_roce_table_put(struct hns_roce_dev *hr_dev,
 
 	mutex_unlock(&table->mutex);
 }
+EXPORT_SYMBOL_GPL(hns_roce_table_put);
 
 void *hns_roce_table_find(struct hns_roce_dev *hr_dev,
 			  struct hns_roce_hem_table *table,
@@ -921,6 +941,22 @@ int hns_roce_init_hem_table(struct hns_roce_dev *hr_dev,
 			num_bt_l0 = hr_dev->caps.sccc_bt_num;
 			hop_num = hr_dev->caps.sccc_hop_num;
 			break;
+		case HEM_TYPE_QPC_TIMER:
+			buf_chunk_size = 1 << (hr_dev->caps.qpc_timer_buf_pg_sz
+					+ PAGE_SHIFT);
+			bt_chunk_size = 1 << (hr_dev->caps.qpc_timer_ba_pg_sz
+					+ PAGE_SHIFT);
+			num_bt_l0 = hr_dev->caps.qpc_timer_bt_num;
+			hop_num = hr_dev->caps.qpc_timer_hop_num;
+			break;
+		case HEM_TYPE_CQC_TIMER:
+			buf_chunk_size = 1 << (hr_dev->caps.cqc_timer_buf_pg_sz
+					+ PAGE_SHIFT);
+			bt_chunk_size = 1 << (hr_dev->caps.cqc_timer_ba_pg_sz
+					+ PAGE_SHIFT);
+			num_bt_l0 = hr_dev->caps.cqc_timer_bt_num;
+			hop_num = hr_dev->caps.cqc_timer_hop_num;
+			break;
 		case HEM_TYPE_SRQC:
 			buf_chunk_size = 1 << (hr_dev->caps.srqc_buf_pg_sz
 					+ PAGE_SHIFT);
@@ -1098,6 +1134,12 @@ void hns_roce_cleanup_hem(struct hns_roce_dev *hr_dev)
 		hns_roce_cleanup_hem_table(hr_dev,
 					   &hr_dev->srq_table.table);
 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
+	if (hr_dev->caps.qpc_timer_entry_sz)
+		hns_roce_cleanup_hem_table(hr_dev,
+					   &hr_dev->qpc_timer_table);
+	if (hr_dev->caps.cqc_timer_entry_sz)
+		hns_roce_cleanup_hem_table(hr_dev,
+					   &hr_dev->cqc_timer_table);
 	if (hr_dev->caps.sccc_entry_sz)
 		hns_roce_cleanup_hem_table(hr_dev,
 					   &hr_dev->qp_table.sccc_table);

drivers/infiniband/hw/hns/hns_roce_hem.h

@@ -45,6 +45,8 @@ enum {
 	HEM_TYPE_CQC,
 	HEM_TYPE_SRQC,
 	HEM_TYPE_SCCC,
+	HEM_TYPE_QPC_TIMER,
+	HEM_TYPE_CQC_TIMER,
 
 	 /* UNMAP HEM */
 	HEM_TYPE_MTT,

drivers/infiniband/hw/hns/hns_roce_hw_v2.c

@@ -1085,6 +1085,41 @@ static int hns_roce_query_pf_resource(struct hns_roce_dev *hr_dev)
 	return 0;
 }
 
+static int hns_roce_query_pf_timer_resource(struct hns_roce_dev *hr_dev)
+{
+	struct hns_roce_pf_timer_res_a *req_a;
+	struct hns_roce_cmq_desc desc[2];
+	int ret, i;
+
+	for (i = 0; i < 2; i++) {
+		hns_roce_cmq_setup_basic_desc(&desc[i],
+					      HNS_ROCE_OPC_QUERY_PF_TIMER_RES,
+					      true);
+
+		if (i == 0)
+			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
+		else
+			desc[i].flag &= ~cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
+	}
+
+	ret = hns_roce_cmq_send(hr_dev, desc, 2);
+	if (ret)
+		return ret;
+
+	req_a = (struct hns_roce_pf_timer_res_a *)desc[0].data;
+
+	hr_dev->caps.qpc_timer_bt_num =
+				roce_get_field(req_a->qpc_timer_bt_idx_num,
+					PF_RES_DATA_1_PF_QPC_TIMER_BT_NUM_M,
+					PF_RES_DATA_1_PF_QPC_TIMER_BT_NUM_S);
+	hr_dev->caps.cqc_timer_bt_num =
+				roce_get_field(req_a->cqc_timer_bt_idx_num,
+					PF_RES_DATA_2_PF_CQC_TIMER_BT_NUM_M,
+					PF_RES_DATA_2_PF_CQC_TIMER_BT_NUM_S);
+
+	return 0;
+}
+
 static int hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev,
 						  int vf_id)
 {
@@ -1315,6 +1350,16 @@ static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
 		return ret;
 	}
 
+	if (hr_dev->pci_dev->revision == 0x21) {
+		ret = hns_roce_query_pf_timer_resource(hr_dev);
+		if (ret) {
+			dev_err(hr_dev->dev,
+				"Query pf timer resource fail, ret = %d.\n",
+				ret);
+			return ret;
+		}
+	}
+
 	ret = hns_roce_alloc_vf_resource(hr_dev);
 	if (ret) {
 		dev_err(hr_dev->dev, "Allocate vf resource fail, ret = %d.\n",
@@ -1439,6 +1484,17 @@ static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
 			       HNS_ROCE_CAP_FLAG_SRQ |
 			       HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL;
 
+		caps->num_qpc_timer	  = HNS_ROCE_V2_MAX_QPC_TIMER_NUM;
+		caps->qpc_timer_entry_sz  = HNS_ROCE_V2_QPC_TIMER_ENTRY_SZ;
+		caps->qpc_timer_ba_pg_sz  = 0;
+		caps->qpc_timer_buf_pg_sz = 0;
+		caps->qpc_timer_hop_num   = HNS_ROCE_HOP_NUM_0;
+		caps->num_cqc_timer	  = HNS_ROCE_V2_MAX_CQC_TIMER_NUM;
+		caps->cqc_timer_entry_sz  = HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ;
+		caps->cqc_timer_ba_pg_sz  = 0;
+		caps->cqc_timer_buf_pg_sz = 0;
+		caps->cqc_timer_hop_num   = HNS_ROCE_HOP_NUM_0;
+
 		caps->sccc_entry_sz	= HNS_ROCE_V2_SCCC_ENTRY_SZ;
 		caps->sccc_ba_pg_sz	= 0;
 		caps->sccc_buf_pg_sz    = 0;
@@ -1644,7 +1700,8 @@ static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev,
 static int hns_roce_v2_init(struct hns_roce_dev *hr_dev)
 {
 	struct hns_roce_v2_priv *priv = hr_dev->priv;
-	int ret;
+	int qpc_count, cqc_count;
+	int ret, i;
 
 	/* TSQ includes SQ doorbell and ack doorbell */
 	ret = hns_roce_init_link_table(hr_dev, TSQ_LINK_TABLE);
@@ -1659,8 +1716,40 @@ static int hns_roce_v2_init(struct hns_roce_dev *hr_dev)
 		goto err_tpq_init_failed;
 	}
 
+	/* Alloc memory for QPC Timer buffer space chunk*/
+	for (qpc_count = 0; qpc_count < hr_dev->caps.qpc_timer_bt_num;
+	     qpc_count++) {
+		ret = hns_roce_table_get(hr_dev, &hr_dev->qpc_timer_table,
+					 qpc_count);
+		if (ret) {
+			dev_err(hr_dev->dev, "QPC Timer get failed\n");
+			goto err_qpc_timer_failed;
+		}
+	}
+
+	/* Alloc memory for CQC Timer buffer space chunk*/
+	for (cqc_count = 0; cqc_count < hr_dev->caps.cqc_timer_bt_num;
+	     cqc_count++) {
+		ret = hns_roce_table_get(hr_dev, &hr_dev->cqc_timer_table,
+					 cqc_count);
+		if (ret) {
+			dev_err(hr_dev->dev, "CQC Timer get failed\n");
+			goto err_cqc_timer_failed;
+		}
+	}
+
 	return 0;
 
+err_cqc_timer_failed:
+	for (i = 0; i < cqc_count; i++)
+		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
+
+err_qpc_timer_failed:
+	for (i = 0; i < qpc_count; i++)
+		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
+
+	hns_roce_free_link_table(hr_dev, &priv->tpq);
+
 err_tpq_init_failed:
 	hns_roce_free_link_table(hr_dev, &priv->tsq);
 
@@ -2699,6 +2788,12 @@ static int hns_roce_v2_set_hem(struct hns_roce_dev *hr_dev,
 	case HEM_TYPE_SCCC:
 		op = HNS_ROCE_CMD_WRITE_SCCC_BT0;
 		break;
+	case HEM_TYPE_QPC_TIMER:
+		op = HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0;
+		break;
+	case HEM_TYPE_CQC_TIMER:
+		op = HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0;
+		break;
 	default:
 		dev_warn(dev, "Table %d not to be written by mailbox!\n",
 			 table->type);
@@ -2763,6 +2858,8 @@ static int hns_roce_v2_clear_hem(struct hns_roce_dev *hr_dev,
 		op = HNS_ROCE_CMD_DESTROY_CQC_BT0;
 		break;
 	case HEM_TYPE_SCCC:
+	case HEM_TYPE_QPC_TIMER:
+	case HEM_TYPE_CQC_TIMER:
 		break;
 	case HEM_TYPE_SRQC:
 		op = HNS_ROCE_CMD_DESTROY_SRQC_BT0;
@@ -2773,7 +2870,9 @@ static int hns_roce_v2_clear_hem(struct hns_roce_dev *hr_dev,
 		return 0;
 	}
 
-	if (table->type == HEM_TYPE_SCCC)
+	if (table->type == HEM_TYPE_SCCC ||
+	    table->type == HEM_TYPE_QPC_TIMER ||
+	    table->type == HEM_TYPE_CQC_TIMER)
 		return 0;
 
 	op += step_idx;

drivers/infiniband/hw/hns/hns_roce_hw_v2.h

@@ -46,11 +46,13 @@
 #define HNS_ROCE_VF_SL_NUM			8
 
 #define HNS_ROCE_V2_MAX_QP_NUM			0x2000
+#define HNS_ROCE_V2_MAX_QPC_TIMER_NUM          0x200
 #define HNS_ROCE_V2_MAX_WQE_NUM			0x8000
 #define	HNS_ROCE_V2_MAX_SRQ			0x100000
 #define HNS_ROCE_V2_MAX_SRQ_WR			0x8000
 #define HNS_ROCE_V2_MAX_SRQ_SGE			0x100
 #define HNS_ROCE_V2_MAX_CQ_NUM			0x8000
+#define HNS_ROCE_V2_MAX_CQC_TIMER_NUM          0x100
 #define HNS_ROCE_V2_MAX_SRQ_NUM			0x100000
 #define HNS_ROCE_V2_MAX_CQE_NUM			0x10000
 #define HNS_ROCE_V2_MAX_SRQWQE_NUM		0x8000
@@ -85,6 +87,8 @@
 #define HNS_ROCE_V2_MTT_ENTRY_SZ		64
 #define HNS_ROCE_V2_CQE_ENTRY_SIZE		32
 #define HNS_ROCE_V2_SCCC_ENTRY_SZ		32
+#define HNS_ROCE_V2_QPC_TIMER_ENTRY_SZ		4096
+#define HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ		4096
 #define HNS_ROCE_V2_PAGE_SIZE_SUPPORTED		0xFFFFF000
 #define HNS_ROCE_V2_MAX_INNER_MTPT_NUM		2
 #define HNS_ROCE_INVALID_LKEY			0x100
@@ -229,6 +233,7 @@ enum hns_roce_opcode_type {
 	HNS_ROCE_OPC_ALLOC_VF_RES			= 0x8401,
 	HNS_ROCE_OPC_CFG_EXT_LLM			= 0x8403,
 	HNS_ROCE_OPC_CFG_TMOUT_LLM			= 0x8404,
+	HNS_ROCE_OPC_QUERY_PF_TIMER_RES			= 0x8406,
 	HNS_ROCE_OPC_CFG_SGID_TB			= 0x8500,
 	HNS_ROCE_OPC_CFG_SMAC_TB			= 0x8501,
 	HNS_ROCE_OPC_POST_MB				= 0x8504,
@@ -1336,6 +1341,25 @@ struct hns_roce_pf_res_b {
 #define PF_RES_DATA_4_PF_SCCC_BT_NUM_S 9
 #define PF_RES_DATA_4_PF_SCCC_BT_NUM_M GENMASK(17, 9)
 
+struct hns_roce_pf_timer_res_a {
+	__le32	rsv0;
+	__le32	qpc_timer_bt_idx_num;
+	__le32	cqc_timer_bt_idx_num;
+	__le32	rsv[3];
+};
+
+#define PF_RES_DATA_1_PF_QPC_TIMER_BT_IDX_S 0
+#define PF_RES_DATA_1_PF_QPC_TIMER_BT_IDX_M GENMASK(11, 0)
+
+#define PF_RES_DATA_1_PF_QPC_TIMER_BT_NUM_S 16
+#define PF_RES_DATA_1_PF_QPC_TIMER_BT_NUM_M GENMASK(28, 16)
+
+#define PF_RES_DATA_2_PF_CQC_TIMER_BT_IDX_S 0
+#define PF_RES_DATA_2_PF_CQC_TIMER_BT_IDX_M GENMASK(10, 0)
+
+#define PF_RES_DATA_2_PF_CQC_TIMER_BT_NUM_S 16
+#define PF_RES_DATA_2_PF_CQC_TIMER_BT_NUM_M GENMASK(27, 16)
+
 struct hns_roce_vf_res_a {
 	__le32 vf_id;
 	__le32 vf_qpc_bt_idx_num;

drivers/infiniband/hw/hns/hns_roce_main.c

@@ -715,8 +715,44 @@ static int hns_roce_init_hem(struct hns_roce_dev *hr_dev)
 		}
 	}
 
+	if (hr_dev->caps.qpc_timer_entry_sz) {
+		ret = hns_roce_init_hem_table(hr_dev,
+					      &hr_dev->qpc_timer_table,
+					      HEM_TYPE_QPC_TIMER,
+					      hr_dev->caps.qpc_timer_entry_sz,
+					      hr_dev->caps.num_qpc_timer, 1);
+		if (ret) {
+			dev_err(dev,
+			      "Failed to init QPC timer memory, aborting.\n");
+			goto err_unmap_ctx;
+		}
+	}
+
+	if (hr_dev->caps.cqc_timer_entry_sz) {
+		ret = hns_roce_init_hem_table(hr_dev,
+					      &hr_dev->cqc_timer_table,
+					      HEM_TYPE_CQC_TIMER,
+					      hr_dev->caps.cqc_timer_entry_sz,
+					      hr_dev->caps.num_cqc_timer, 1);
+		if (ret) {
+			dev_err(dev,
+			      "Failed to init CQC timer memory, aborting.\n");
+			goto err_unmap_qpc_timer;
+		}
+	}
+
 	return 0;
 
+err_unmap_qpc_timer:
+	if (hr_dev->caps.qpc_timer_entry_sz)
+		hns_roce_cleanup_hem_table(hr_dev,
+					   &hr_dev->qpc_timer_table);
+
+err_unmap_ctx:
+	if (hr_dev->caps.sccc_entry_sz)
+		hns_roce_cleanup_hem_table(hr_dev,
+					   &hr_dev->qp_table.sccc_table);
+
 err_unmap_idx:
 	if (hr_dev->caps.num_idx_segs)
 		hns_roce_cleanup_hem_table(hr_dev,