diff mbox

[v5,08/21] IB/hns: Add icm support

Message ID 1461407219-72027-9-git-send-email-oulijun@huawei.com (mailing list archive)
State Superseded
Headers show

Commit Message

Lijun Ou April 23, 2016, 10:26 a.m. UTC
This patch mainly added icm support for RoCE. It inits the icm
which managers the relative memory blocks for RoCE. the data
structures of RoCE will be located in it. for example, CQ table,
QP table and MTPT table so on.

Signed-off-by: Lijun Ou <oulijun@huawei.com>
Signed-off-by: Wei Hu(Xavier) <xavier.huwei@huawei.com>
---
 drivers/infiniband/hw/hns/hns_roce_common.h |  19 ++
 drivers/infiniband/hw/hns/hns_roce_device.h |  30 ++
 drivers/infiniband/hw/hns/hns_roce_eq.c     |   3 +-
 drivers/infiniband/hw/hns/hns_roce_icm.c    | 436 ++++++++++++++++++++++++++++
 drivers/infiniband/hw/hns/hns_roce_icm.h    |  95 ++++++
 drivers/infiniband/hw/hns/hns_roce_main.c   |  84 ++++++
 6 files changed, 665 insertions(+), 2 deletions(-)
 create mode 100644 drivers/infiniband/hw/hns/hns_roce_icm.c
 create mode 100644 drivers/infiniband/hw/hns/hns_roce_icm.h
diff mbox

Patch

diff --git a/drivers/infiniband/hw/hns/hns_roce_common.h b/drivers/infiniband/hw/hns/hns_roce_common.h
index 2d083c2..5e181de 100644
--- a/drivers/infiniband/hw/hns/hns_roce_common.h
+++ b/drivers/infiniband/hw/hns/hns_roce_common.h
@@ -30,6 +30,22 @@ 
 #define roce_set_bit(origin, shift, val) \
 	roce_set_field((origin), (1ul << (shift)), (shift), (val))
 
+#define ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S 0
+#define ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_M   \
+	(((1UL << 19) - 1) << ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S)
+
+#define ROCEE_BT_CMD_H_ROCEE_BT_CMD_S 19
+
+#define ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S 20
+#define ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M   \
+	(((1UL << 2) - 1) << ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S)
+
+#define ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S 22
+#define ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_M   \
+	(((1UL << 5) - 1) << ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S)
+
+#define ROCEE_BT_CMD_H_ROCEE_BT_CMD_HW_SYNS_S 31
+
 #define ROCEE_CAEP_AEQC_AEQE_SHIFT_CAEP_AEQC_STATE_S 0
 #define ROCEE_CAEP_AEQC_AEQE_SHIFT_CAEP_AEQC_STATE_M   \
 	(((1UL << 2) - 1) << ROCEE_CAEP_AEQC_AEQE_SHIFT_CAEP_AEQC_STATE_S)
@@ -70,6 +86,8 @@ 
 #define ROCEE_SYS_IMAGE_GUID_L_REG		0xC
 #define ROCEE_SYS_IMAGE_GUID_H_REG		0x10
 
+#define ROCEE_BT_CMD_H_REG			0x204
+
 #define ROCEE_CAEP_AEQE_CONS_IDX_REG		0x3AC
 #define ROCEE_CAEP_CEQC_CONS_IDX_0_REG		0x3BC
 
@@ -82,6 +100,7 @@ 
 
 #define ROCEE_CAEP_CE_INTERVAL_CFG_REG		0x190
 #define ROCEE_CAEP_CE_BURST_NUM_CFG_REG		0x194
+#define ROCEE_BT_CMD_L_REG			0x200
 
 #define ROCEE_MB1_REG				0x210
 
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index 6835223..decd4fe 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -91,6 +91,26 @@  enum {
 	HNS_ROCE_CMD_SUCCESS			= 1,
 };
 
+struct hns_roce_icm_table {
+	/* ICM type: 0 = qpc 1 = mtt 2 = cqc 3 = srq 4 = other */
+	u32            type;
+	/* ICM array elment num */
+	int            num_icm;
+	/* ICM entry record obj total num */
+	int            num_obj;
+	/*Single obj size */
+	int            obj_size;
+	int            lowmem;
+	int            coherent;
+	struct mutex   mutex;
+	struct hns_roce_icm **icm;
+};
+
+struct hns_roce_mr_table {
+	struct hns_roce_icm_table	mtt_table;
+	struct hns_roce_icm_table	mtpt_table;
+};
+
 struct hns_roce_buf_list {
 	void		*buf;
 	dma_addr_t	map;
@@ -106,11 +126,14 @@  struct hns_roce_cq {
 
 struct hns_roce_qp_table {
 	spinlock_t			lock;
+	struct hns_roce_icm_table	qp_table;
+	struct hns_roce_icm_table	irrl_table;
 };
 
 struct hns_roce_cq_table {
 	spinlock_t			lock;
 	struct radix_tree_root		tree;
+	struct hns_roce_icm_table	table;
 };
 
 struct hns_roce_cmd_context {
@@ -239,6 +262,7 @@  struct hns_roce_hw {
 struct hns_roce_dev {
 	struct ib_device	ib_dev;
 	struct platform_device  *pdev;
+	spinlock_t		bt_cmd_lock;
 	struct hns_roce_ib_iboe iboe;
 
 	int			irq[HNS_ROCE_MAX_IRQ_NUM];
@@ -253,6 +277,7 @@  struct hns_roce_dev {
 	u32                     hw_rev;
 
 	struct hns_roce_cmdq      cmd;
+	struct hns_roce_mr_table  mr_table;
 	struct hns_roce_cq_table  cq_table;
 	struct hns_roce_qp_table  qp_table;
 	struct hns_roce_eq_table  eq_table;
@@ -262,6 +287,11 @@  struct hns_roce_dev {
 	struct hns_roce_hw	*hw;
 };
 
+static inline void hns_roce_write64_k(__be32 val[2], void __iomem *dest)
+{
+	__raw_writeq(*(u64 *) val, dest);
+}
+
 static inline struct hns_roce_qp
 	*__hns_roce_qp_lookup(struct hns_roce_dev *hr_dev, u32 qpn)
 {
diff --git a/drivers/infiniband/hw/hns/hns_roce_eq.c b/drivers/infiniband/hw/hns/hns_roce_eq.c
index d8ca66f..7d848a0 100644
--- a/drivers/infiniband/hw/hns/hns_roce_eq.c
+++ b/drivers/infiniband/hw/hns/hns_roce_eq.c
@@ -589,8 +589,7 @@  void hns_roce_int_mask_en(struct hns_roce_dev *hr_dev)
 	int masken = 0;
 
 	/* AEQ INT */
-	reg_caepaemsk = (void *)((char *)hr_dev->reg_base +
-				  ROCEE_CAEP_AE_MASK_REG);
+	reg_caepaemsk = (void *)(hr_dev->reg_base + ROCEE_CAEP_AE_MASK_REG);
 	aemask_val = roce_readl(reg_caepaemsk);
 	roce_set_bit(aemask_val, ROCEE_CAEP_AE_MASK_CAEP_AEQ_ALM_OVF_MASK_S,
 		     masken);
diff --git a/drivers/infiniband/hw/hns/hns_roce_icm.c b/drivers/infiniband/hw/hns/hns_roce_icm.c
new file mode 100644
index 0000000..d0f276d
--- /dev/null
+++ b/drivers/infiniband/hw/hns/hns_roce_icm.c
@@ -0,0 +1,436 @@ 
+/*
+ * Copyright (c) 2016 Hisilicon Limited.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ */
+
+#include <linux/errno.h>
+#include <linux/init.h>
+#include <linux/mm.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+#include <linux/scatterlist.h>
+#include "hns_roce_device.h"
+#include "hns_roce_icm.h"
+#include "hns_roce_common.h"
+
+#define HW_SYNC_TIMEOUT_MSECS		500
+
+#define HNS_ROCE_ICM_ALLOC_SIZE		(1 << 17)
+#define HNS_ROCE_TABLE_CHUNK_SIZE	(1 << 17)
+
+#define DMA_ADDR_T_SHIFT		12
+#define BT_CMD_SYNC_SHIFT		31
+#define BT_BA_SHIFT			32
+
+static int hns_roce_alloc_icm_pages(struct scatterlist *mem, int order,
+				    gfp_t gfp_mask)
+{
+	struct page *page;
+
+	page = alloc_pages(gfp_mask, order);
+	if (!page)
+		return -ENOMEM;
+
+	sg_set_page(mem, page, PAGE_SIZE << order, 0);
+
+	return 0;
+}
+
+static int hns_roce_alloc_icm_coherent(struct device *dev,
+				       struct scatterlist *mem, int order,
+				       gfp_t gfp_mask)
+{
+	void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order,
+				       &sg_dma_address(mem), gfp_mask);
+	if (!buf) {
+		dev_err(dev, "alloc coherent icm pages failed.\n");
+		return -ENOMEM;
+	}
+
+	sg_set_buf(mem, buf, PAGE_SIZE << order);
+	WARN_ON(mem->offset);
+	sg_dma_len(mem) = PAGE_SIZE << order;
+	return 0;
+}
+
+struct hns_roce_icm *hns_roce_alloc_icm(struct hns_roce_dev *hr_dev, int npages,
+					gfp_t gfp_mask, int coherent)
+{
+	struct hns_roce_icm_chunk *chunk = NULL;
+	struct hns_roce_icm *icm;
+	int cur_order;
+	int ret;
+
+	WARN_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
+
+	icm = kmalloc(sizeof(*icm),
+		      gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
+	if (!icm)
+		return NULL;
+
+	icm->refcount = 0;
+	INIT_LIST_HEAD(&icm->chunk_list);
+
+	cur_order = get_order(HNS_ROCE_ICM_ALLOC_SIZE);
+
+	while (npages > 0) {
+		if (!chunk) {
+			chunk = kmalloc(sizeof(*chunk),
+				gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
+			if (!chunk)
+				goto fail;
+
+			sg_init_table(chunk->mem, HNS_ROCE_ICM_CHUNK_LEN);
+			chunk->npages = 0;
+			chunk->nsg = 0;
+			list_add_tail(&chunk->list, &icm->chunk_list);
+		}
+
+		while (1 << cur_order > npages)
+			--cur_order;
+
+		if (coherent)
+			ret = hns_roce_alloc_icm_coherent(&hr_dev->pdev->dev,
+						&chunk->mem[chunk->npages],
+						cur_order, gfp_mask);
+		else
+			ret = hns_roce_alloc_icm_pages(
+						&chunk->mem[chunk->npages],
+						cur_order, gfp_mask);
+		if (!ret) {
+			++chunk->npages;
+			if (coherent) {
+				++chunk->nsg;
+			} else if (chunk->npages == HNS_ROCE_ICM_CHUNK_LEN) {
+				chunk->nsg = dma_map_sg(&hr_dev->pdev->dev,
+						      chunk->mem, chunk->npages,
+						      DMA_BIDIRECTIONAL);
+				if (chunk->nsg <= 0)
+					goto fail;
+
+				chunk = NULL;
+			}
+			npages -= 1 << cur_order;
+		} else {
+			/*
+			* If failed on alloc 128k memory one time,
+			* no alloc small block memory,
+			* directly return fail
+			*/
+			dev_err(&hr_dev->pdev->dev, "**alloc 128K phy mem failed**\n");
+			goto fail;
+		}
+	}
+
+	if (!coherent && chunk) {
+		chunk->nsg = dma_map_sg(&hr_dev->pdev->dev, chunk->mem,
+					chunk->npages, DMA_BIDIRECTIONAL);
+		if (chunk->nsg <= 0)
+			goto fail;
+	}
+
+	return icm;
+
+fail:
+	hns_roce_free_icm(hr_dev, icm, coherent);
+	return NULL;
+}
+
+static void hns_roce_free_icm_pages(struct hns_roce_dev *hr_dev,
+				    struct hns_roce_icm_chunk *chunk)
+{
+	int i;
+
+	if (chunk->nsg > 0)
+		dma_unmap_sg(&hr_dev->pdev->dev, chunk->mem, chunk->npages,
+			     DMA_BIDIRECTIONAL);
+
+	for (i = 0; i < chunk->npages; ++i)
+		__free_pages(sg_page(&chunk->mem[i]),
+			     get_order(chunk->mem[i].length));
+}
+
+static void hns_roce_free_icm_coherent(struct hns_roce_dev *hr_dev,
+				       struct hns_roce_icm_chunk *chunk)
+{
+	int i;
+
+	for (i = 0; i < chunk->npages; ++i)
+		dma_free_coherent(&hr_dev->pdev->dev, chunk->mem[i].length,
+				  lowmem_page_address(sg_page(&chunk->mem[i])),
+				  sg_dma_address(&chunk->mem[i]));
+}
+
+void hns_roce_free_icm(struct hns_roce_dev *hr_dev, struct hns_roce_icm *icm,
+		       int coherent)
+{
+	struct hns_roce_icm_chunk *chunk, *tmp;
+
+	if (!icm)
+		return;
+
+	list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
+		if (coherent)
+			hns_roce_free_icm_coherent(hr_dev, chunk);
+		else
+			hns_roce_free_icm_pages(hr_dev, chunk);
+
+		kfree(chunk);
+	}
+
+	kfree(icm);
+}
+
+int hns_roce_map_icm(struct hns_roce_dev *hr_dev,
+		     struct hns_roce_icm_table *table, int obj)
+{
+	struct device *dev = &hr_dev->pdev->dev;
+	spinlock_t *lock = &hr_dev->bt_cmd_lock;
+	unsigned long end = 0, flags = 0;
+	struct hns_roce_icm_iter iter;
+	void __iomem *bt_cmd = NULL;
+	u32 bt_cmd_h_val = 0;
+	u32 bt_cmd_val[2];
+	u32 bt_cmd_l = 0;
+	u64 bt_ba = 0;
+	int ret = 0;
+
+	/* Find the icm entry */
+	int i = (obj & (table->num_obj - 1)) / (HNS_ROCE_TABLE_CHUNK_SIZE /
+		table->obj_size);
+
+	switch (table->type) {
+	case ICM_TYPE_QPC:
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, ICM_TYPE_QPC);
+		break;
+	case ICM_TYPE_MTPT:
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S,
+			       ICM_TYPE_MTPT);
+		break;
+	case ICM_TYPE_CQC:
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, ICM_TYPE_CQC);
+		break;
+	case ICM_TYPE_SRQC:
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S,
+			       ICM_TYPE_SRQC);
+		break;
+	default:
+		return ret;
+	}
+	roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_M,
+		       ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S, obj);
+	roce_set_bit(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_S, 0);
+	roce_set_bit(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_HW_SYNS_S, 1);
+
+	/* Currently iter only a chunk */
+	for (hns_roce_icm_first(table->icm[i], &iter);
+	     !hns_roce_icm_last(&iter); hns_roce_icm_next(&iter)) {
+		bt_ba = hns_roce_icm_addr(&iter) >> DMA_ADDR_T_SHIFT;
+
+		spin_lock_irqsave(lock, flags);
+
+		bt_cmd = hr_dev->reg_base + ROCEE_BT_CMD_H_REG;
+
+		end = msecs_to_jiffies(HW_SYNC_TIMEOUT_MSECS) + jiffies;
+		while (1) {
+			if (roce_readl(bt_cmd) >> BT_CMD_SYNC_SHIFT) {
+				if (!(time_before(jiffies, end))) {
+					dev_err(dev, "Write bt_cmd err,hw_sync is not zero.\n");
+					spin_unlock_irqrestore(lock, flags);
+					ret = -EBUSY;
+					return ret;
+				}
+			} else {
+				break;
+			}
+		}
+
+		bt_cmd_l = (u32)bt_ba;
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S,
+			       bt_ba >> BT_BA_SHIFT);
+
+		bt_cmd_val[0] = bt_cmd_l;
+		bt_cmd_val[1] = bt_cmd_h_val;
+		hns_roce_write64_k(bt_cmd_val,
+				   hr_dev->reg_base + ROCEE_BT_CMD_L_REG);
+		spin_unlock_irqrestore(lock, flags);
+	}
+
+	return ret;
+}
+
+int hns_roce_unmap_icm(struct hns_roce_dev *hr_dev,
+		       struct hns_roce_icm_table *table, int obj)
+{
+	struct device *dev = &hr_dev->pdev->dev;
+	unsigned long end = 0;
+	unsigned long flag;
+	void __iomem *bt_cmd;
+	uint32_t bt_cmd_val[2];
+	u32 bt_cmd_h_val = 0;
+	int ret = 0;
+
+	switch (table->type) {
+	case ICM_TYPE_QPC:
+		dev_dbg(dev, "UNMAP QPC BT  :\n");
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, ICM_TYPE_QPC);
+		break;
+	case ICM_TYPE_MTPT:
+		dev_dbg(dev, "UNMAP MTPT BT :\n");
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S,
+			       ICM_TYPE_MTPT);
+		break;
+	case ICM_TYPE_CQC:
+		dev_dbg(dev, "UNMAP CQC BT  :\n");
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, ICM_TYPE_CQC);
+		break;
+	case ICM_TYPE_SRQC:
+		roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S,
+			       ICM_TYPE_SRQC);
+		break;
+	default:
+		return ret;
+	}
+	roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_M,
+		       ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S, obj);
+	roce_set_bit(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_S, 1);
+	roce_set_bit(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_HW_SYNS_S, 1);
+	roce_set_field(bt_cmd_h_val, ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_M,
+		       ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S, 0);
+
+	spin_lock_irqsave(&hr_dev->bt_cmd_lock, flags);
+
+	bt_cmd = hr_dev->reg_base + ROCEE_BT_CMD_H_REG;
+
+	end = msecs_to_jiffies(HW_SYNC_TIMEOUT_MSECS) + jiffies;
+	while (1) {
+		msleep(20);
+		if (roce_readl(bt_cmd) >> BT_CMD_SYNC_SHIFT) {
+			if (!(time_before(jiffies, end))) {
+				dev_err(dev, "Write bt_cmd err,hw_sync is not zero.\n");
+				spin_unlock_irqrestore(&hr_dev->bt_cmd_lock,
+						       flags);
+				return -EBUSY;
+			}
+		} else {
+			break;
+		}
+	}
+
+	bt_cmd_val[0] = 0;
+	bt_cmd_val[1] = bt_cmd_h_val;
+	hns_roce_write64_k(bt_cmd_val, hr_dev->reg_base + ROCEE_BT_CMD_L_REG);
+	spin_unlock_irqrestore(&hr_dev->bt_cmd_lock, flags);
+
+	return ret;
+}
+
+int hns_roce_init_icm_table(struct hns_roce_dev *hr_dev,
+			    struct hns_roce_icm_table *table, u32 type,
+			    int obj_size, int nobj, int reserved,
+			    int use_lowmem, int use_coherent)
+{
+	struct device *dev = &hr_dev->pdev->dev;
+	u32 chunk_size;
+	int obj_per_chunk;
+	int num_icm;
+	int i;
+
+	obj_per_chunk = HNS_ROCE_TABLE_CHUNK_SIZE / obj_size;
+	num_icm = (nobj + obj_per_chunk - 1) / obj_per_chunk;
+
+	table->icm = kcalloc(num_icm, sizeof(*table->icm), GFP_KERNEL);
+	if (!table->icm)
+		return -ENOMEM;
+
+	table->type = type;
+	table->num_icm = num_icm;
+	table->num_obj = nobj;
+	table->obj_size = obj_size;
+	table->lowmem = use_lowmem;
+	table->coherent = use_coherent;
+	mutex_init(&table->mutex);
+
+	for (i = 0;
+	i * HNS_ROCE_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
+		chunk_size = HNS_ROCE_TABLE_CHUNK_SIZE;
+		if ((i + 1) * HNS_ROCE_TABLE_CHUNK_SIZE > nobj * obj_size)
+			chunk_size = PAGE_ALIGN(nobj * obj_size -
+						i * HNS_ROCE_TABLE_CHUNK_SIZE);
+
+		table->icm[i] = hns_roce_alloc_icm(hr_dev,
+				chunk_size >> PAGE_SHIFT, (use_lowmem ?
+				GFP_KERNEL : GFP_HIGHUSER) | __GFP_NOWARN,
+				use_coherent);
+		if (!table->icm[i])
+			goto _error_failed_alloc_icm;
+
+		if (hns_roce_map_icm(hr_dev, table,
+			i * HNS_ROCE_TABLE_CHUNK_SIZE / obj_size)) {
+			dev_err(dev, "map icm table failed.\n");
+			goto _error_failed_map_icm;
+		}
+
+		/*
+		 * Add a reference to this ICM chunk so that it never
+		 * Gets freed (since it contains reserved firmware objects).
+		 */
+		++table->icm[i]->refcount;
+	}
+
+	return 0;
+
+_error_failed_map_icm:
+_error_failed_alloc_icm:
+	for (i = 0; i < num_icm; ++i)
+		if (table->icm[i]) {
+			if (hns_roce_unmap_icm(hr_dev, table,
+				i * HNS_ROCE_TABLE_CHUNK_SIZE / obj_size))
+				dev_err(dev, "unmap icm table failed.\n");
+
+			hns_roce_free_icm(hr_dev, table->icm[i], use_coherent);
+		}
+
+	return -ENOMEM;
+}
+
+void hns_roce_cleanup_icm_table(struct hns_roce_dev *hr_dev,
+				struct hns_roce_icm_table *table)
+{
+	struct device *dev = &hr_dev->pdev->dev;
+	int i;
+
+	for (i = 0; i < table->num_icm; ++i)
+		if (table->icm[i]) {
+			if (hns_roce_unmap_icm(hr_dev, table,
+			    i * HNS_ROCE_TABLE_CHUNK_SIZE / table->obj_size))
+				dev_err(dev, "unmap icm table failed.\n");
+
+			hns_roce_free_icm(hr_dev, table->icm[i],
+					  table->coherent);
+		}
+
+	kfree(table->icm);
+}
+
+void hns_roce_cleanup_icm(struct hns_roce_dev *hr_dev)
+{
+	hns_roce_cleanup_icm_table(hr_dev, &hr_dev->cq_table.table);
+	hns_roce_cleanup_icm_table(hr_dev, &hr_dev->qp_table.irrl_table);
+	hns_roce_cleanup_icm_table(hr_dev, &hr_dev->qp_table.qp_table);
+	hns_roce_cleanup_icm_table(hr_dev, &hr_dev->mr_table.mtpt_table);
+	hns_roce_cleanup_icm_table(hr_dev, &hr_dev->mr_table.mtt_table);
+}
diff --git a/drivers/infiniband/hw/hns/hns_roce_icm.h b/drivers/infiniband/hw/hns/hns_roce_icm.h
new file mode 100644
index 0000000..ce96d89
--- /dev/null
+++ b/drivers/infiniband/hw/hns/hns_roce_icm.h
@@ -0,0 +1,95 @@ 
+/*
+ * Copyright (c) 2016 Hisilicon Limited.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ */
+
+#ifndef _HNS_ROCE_ICM_H
+#define _HNS_ROCE_ICM_H
+
+#include <linux/list.h>
+#include <linux/mutex.h>
+#include <linux/pci.h>
+
+enum {
+	/* MAP ICM */
+	ICM_TYPE_QPC = 0,
+	ICM_TYPE_MTPT,
+	ICM_TYPE_CQC,
+	ICM_TYPE_SRQC,
+
+	 /* UNMAP ICM */
+	ICM_TYPE_MTT,
+	ICM_TYPE_IRRL,
+};
+
+#define HNS_ROCE_ICM_CHUNK_LEN	\
+	 ((256 - sizeof(struct list_head) - 2 * sizeof(int)) /	 \
+	 (sizeof(struct scatterlist)))
+
+struct hns_roce_icm_chunk {
+	struct list_head	 list;
+	int			 npages;
+	int			 nsg;
+	struct scatterlist	 mem[HNS_ROCE_ICM_CHUNK_LEN];
+};
+
+struct hns_roce_icm {
+	struct list_head	 chunk_list;
+	int			 refcount;
+};
+
+struct hns_roce_icm_iter {
+	struct hns_roce_icm		 *icm;
+	struct hns_roce_icm_chunk	 *chunk;
+	int				 page_idx;
+};
+
+void hns_roce_free_icm(struct hns_roce_dev *hr_dev,
+		       struct hns_roce_icm *icm, int coherent);
+int hns_roce_init_icm_table(struct hns_roce_dev *hr_dev,
+			    struct hns_roce_icm_table *table, u32 type,
+			    int obj_size, int nobj, int reserved,
+			    int use_lowmem, int use_coherent);
+void hns_roce_cleanup_icm_table(struct hns_roce_dev *hr_dev,
+				struct hns_roce_icm_table *table);
+void hns_roce_cleanup_icm(struct hns_roce_dev *hr_dev);
+
+static inline void hns_roce_icm_first(struct hns_roce_icm *icm,
+				      struct hns_roce_icm_iter *iter)
+{
+	iter->icm = icm;
+	iter->chunk = list_empty(&icm->chunk_list) ? NULL :
+				 list_entry(icm->chunk_list.next,
+					    struct hns_roce_icm_chunk, list);
+	 iter->page_idx = 0;
+}
+
+static inline int hns_roce_icm_last(struct hns_roce_icm_iter *iter)
+{
+	return !iter->chunk;
+}
+
+static inline void hns_roce_icm_next(struct hns_roce_icm_iter *iter)
+{
+	if (++iter->page_idx >= iter->chunk->nsg) {
+		if (iter->chunk->list.next == &iter->icm->chunk_list) {
+			iter->chunk = NULL;
+			return;
+		}
+
+		iter->chunk = list_entry(iter->chunk->list.next,
+					 struct hns_roce_icm_chunk, list);
+		iter->page_idx = 0;
+	}
+}
+
+static inline dma_addr_t hns_roce_icm_addr(struct hns_roce_icm_iter *iter)
+{
+	return sg_dma_address(&iter->chunk->mem[iter->page_idx]);
+}
+
+#endif /*_HNS_ROCE_ICM_H*/
diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c
index ea10fb5..3da0c03 100644
--- a/drivers/infiniband/hw/hns/hns_roce_main.c
+++ b/drivers/infiniband/hw/hns/hns_roce_main.c
@@ -41,6 +41,7 @@ 
 #include <rdma/ib_verbs.h>
 #include "hns_roce_common.h"
 #include "hns_roce_device.h"
+#include "hns_roce_icm.h"
 
 int hns_roce_get_cfg(struct hns_roce_dev *hr_dev)
 {
@@ -111,6 +112,77 @@  void hns_roce_profile_init(struct hns_roce_dev *hr_dev)
 	hr_dev->hw->hw_profile(hr_dev);
 }
 
+int hns_roce_init_icm(struct hns_roce_dev *hr_dev)
+{
+	int ret;
+	struct device *dev = &hr_dev->pdev->dev;
+
+	ret = hns_roce_init_icm_table(hr_dev,
+				      (void *)&hr_dev->mr_table.mtt_table,
+				      ICM_TYPE_MTT, hr_dev->caps.mtt_entry_sz,
+				      hr_dev->caps.num_mtt_segs, 0, 1, 0);
+	if (ret) {
+		dev_err(dev, "Failed to map MTT context memory, aborting.\n");
+		return ret;
+	}
+
+	ret = hns_roce_init_icm_table(hr_dev,
+				      (void *)&hr_dev->mr_table.mtpt_table,
+				      ICM_TYPE_MTPT, hr_dev->caps.mtpt_entry_sz,
+				      hr_dev->caps.num_mtpts, 0, 1, 1);
+	if (ret) {
+		dev_err(dev, "Failed to map dMPT context memory, aborting.\n");
+		goto err_unmap_mtt;
+	}
+
+	ret = hns_roce_init_icm_table(hr_dev,
+				      (void *)&hr_dev->qp_table.qp_table,
+				      ICM_TYPE_QPC, hr_dev->caps.qpc_entry_sz,
+				      hr_dev->caps.num_qps, 0, 1, 0);
+	if (ret) {
+		dev_err(dev, "Failed to map QP context memory, aborting.\n");
+		goto err_unmap_dmpt;
+	}
+
+	ret = hns_roce_init_icm_table(hr_dev,
+				      (void *)&hr_dev->qp_table.irrl_table,
+				      ICM_TYPE_IRRL,
+				      hr_dev->caps.irrl_entry_sz *
+				      hr_dev->caps.max_qp_init_rdma,
+				      hr_dev->caps.num_qps, 0, 1, 0);
+	if (ret) {
+		dev_err(dev, "Failed to map irrl_table memory, aborting.\n");
+		goto err_unmap_qp;
+	}
+
+	ret = hns_roce_init_icm_table(hr_dev,
+				      (void *)&hr_dev->cq_table.table,
+				      ICM_TYPE_CQC, hr_dev->caps.cqc_entry_sz,
+				      hr_dev->caps.num_cqs, 0, 1, 0);
+	if (ret) {
+		dev_err(dev, "Failed to map CQ context memory, aborting.\n");
+		goto err_unmap_irrl;
+	}
+
+	return 0;
+
+err_unmap_irrl:
+	hns_roce_cleanup_icm_table(hr_dev,
+				   (void *)&hr_dev->qp_table.irrl_table);
+
+err_unmap_qp:
+	hns_roce_cleanup_icm_table(hr_dev, (void *)&hr_dev->qp_table.qp_table);
+
+err_unmap_dmpt:
+	hns_roce_cleanup_icm_table(hr_dev,
+				   (void *)&hr_dev->mr_table.mtpt_table);
+
+err_unmap_mtt:
+	hns_roce_cleanup_icm_table(hr_dev, (void *)&hr_dev->mr_table.mtt_table);
+
+	return ret;
+}
+
 /**
 * hns_roce_probe - RoCE driver entrance
 * @pdev: pointer to platform device
@@ -178,6 +250,16 @@  static int hns_roce_probe(struct platform_device *pdev)
 		}
 	}
 
+	ret = hns_roce_init_icm(hr_dev);
+	if (ret) {
+		dev_err(dev, "init icm fail!\n");
+		goto error_failed_init_icm;
+	}
+
+error_failed_init_icm:
+	if (hr_dev->cmd_mod)
+		hns_roce_cmd_use_polling(hr_dev);
+
 error_failed_use_event:
 	hns_roce_cleanup_eq_table(hr_dev);
 
@@ -204,6 +286,8 @@  static int hns_roce_remove(struct platform_device *pdev)
 	struct hns_roce_dev *hr_dev = platform_get_drvdata(pdev);
 	int ret = 0;
 
+	hns_roce_cleanup_icm(hr_dev);
+
 	if (hr_dev->cmd_mod)
 		hns_roce_cmd_use_polling(hr_dev);