diff --git a/drivers/staging/Kconfig b/drivers/staging/Kconfig
index 331bc32..8b7b66d 100644
--- a/drivers/staging/Kconfig
+++ b/drivers/staging/Kconfig
@@ -124,4 +124,6 @@
 
 source "drivers/staging/mt7621-spi/Kconfig"
 
+source "drivers/staging/mt7621-dma/Kconfig"
+
 endif # STAGING
diff --git a/drivers/staging/Makefile b/drivers/staging/Makefile
index 6979116..89caabb 100644
--- a/drivers/staging/Makefile
+++ b/drivers/staging/Makefile
@@ -53,3 +53,4 @@
 obj-$(CONFIG_SOC_MT7621)	+= mt7621-pinctrl/
 obj-$(CONFIG_SOC_MT7621)	+= mt7621-gpio/
 obj-$(CONFIG_SOC_MT7621)	+= mt7621-spi/
+obj-$(CONFIG_SOC_MT7621)	+= mt7621-dma/
diff --git a/drivers/staging/mt7621-dma/Kconfig b/drivers/staging/mt7621-dma/Kconfig
new file mode 100644
index 0000000..2423c40
--- /dev/null
+++ b/drivers/staging/mt7621-dma/Kconfig
@@ -0,0 +1,12 @@
+config DMA_RALINK
+	tristate "RALINK DMA support"
+	depends on RALINK && !SOC_RT288X
+	select DMA_ENGINE
+	select DMA_VIRTUAL_CHANNELS
+
+config MTK_HSDMA
+	tristate "MTK HSDMA support"
+	depends on RALINK && SOC_MT7621
+	select DMA_ENGINE
+	select DMA_VIRTUAL_CHANNELS
+
diff --git a/drivers/staging/mt7621-dma/Makefile b/drivers/staging/mt7621-dma/Makefile
new file mode 100644
index 0000000..d3152d4
--- /dev/null
+++ b/drivers/staging/mt7621-dma/Makefile
@@ -0,0 +1,4 @@
+obj-$(CONFIG_DMA_RALINK) += ralink-gdma.o
+obj-$(CONFIG_MTK_HSDMA) += mtk-hsdma.o
+
+ccflags-y += -I$(srctree)/drivers/dma
diff --git a/drivers/staging/mt7621-dma/TODO b/drivers/staging/mt7621-dma/TODO
new file mode 100644
index 0000000..fdbc500
--- /dev/null
+++ b/drivers/staging/mt7621-dma/TODO
@@ -0,0 +1,5 @@
+
+- general code review and clean up
+- ensure device-tree requirements are documented
+
+Cc: NeilBrown <neil@brown.name>
diff --git a/drivers/staging/mt7621-dma/mtk-hsdma.c b/drivers/staging/mt7621-dma/mtk-hsdma.c
new file mode 100644
index 0000000..5eba86e
--- /dev/null
+++ b/drivers/staging/mt7621-dma/mtk-hsdma.c
@@ -0,0 +1,767 @@
+/*
+ *  Copyright (C) 2015, Michael Lee <igvtee@gmail.com>
+ *  MTK HSDMA support
+ *
+ *  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/dmaengine.h>
+#include <linux/dma-mapping.h>
+#include <linux/err.h>
+#include <linux/init.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/irq.h>
+#include <linux/of_dma.h>
+#include <linux/reset.h>
+#include <linux/of_device.h>
+
+#include "virt-dma.h"
+
+#define HSDMA_BASE_OFFSET		0x800
+
+#define HSDMA_REG_TX_BASE		0x00
+#define HSDMA_REG_TX_CNT		0x04
+#define HSDMA_REG_TX_CTX		0x08
+#define HSDMA_REG_TX_DTX		0x0c
+#define HSDMA_REG_RX_BASE		0x100
+#define HSDMA_REG_RX_CNT		0x104
+#define HSDMA_REG_RX_CRX		0x108
+#define HSDMA_REG_RX_DRX		0x10c
+#define HSDMA_REG_INFO			0x200
+#define HSDMA_REG_GLO_CFG		0x204
+#define HSDMA_REG_RST_CFG		0x208
+#define HSDMA_REG_DELAY_INT		0x20c
+#define HSDMA_REG_FREEQ_THRES		0x210
+#define HSDMA_REG_INT_STATUS		0x220
+#define HSDMA_REG_INT_MASK		0x228
+#define HSDMA_REG_SCH_Q01		0x280
+#define HSDMA_REG_SCH_Q23		0x284
+
+#define HSDMA_DESCS_MAX			0xfff
+#define HSDMA_DESCS_NUM			8
+#define HSDMA_DESCS_MASK		(HSDMA_DESCS_NUM - 1)
+#define HSDMA_NEXT_DESC(x)		(((x) + 1) & HSDMA_DESCS_MASK)
+
+/* HSDMA_REG_INFO */
+#define HSDMA_INFO_INDEX_MASK		0xf
+#define HSDMA_INFO_INDEX_SHIFT		24
+#define HSDMA_INFO_BASE_MASK		0xff
+#define HSDMA_INFO_BASE_SHIFT		16
+#define HSDMA_INFO_RX_MASK		0xff
+#define HSDMA_INFO_RX_SHIFT		8
+#define HSDMA_INFO_TX_MASK		0xff
+#define HSDMA_INFO_TX_SHIFT		0
+
+/* HSDMA_REG_GLO_CFG */
+#define HSDMA_GLO_TX_2B_OFFSET		BIT(31)
+#define HSDMA_GLO_CLK_GATE		BIT(30)
+#define HSDMA_GLO_BYTE_SWAP		BIT(29)
+#define HSDMA_GLO_MULTI_DMA		BIT(10)
+#define HSDMA_GLO_TWO_BUF		BIT(9)
+#define HSDMA_GLO_32B_DESC		BIT(8)
+#define HSDMA_GLO_BIG_ENDIAN		BIT(7)
+#define HSDMA_GLO_TX_DONE		BIT(6)
+#define HSDMA_GLO_BT_MASK		0x3
+#define HSDMA_GLO_BT_SHIFT		4
+#define HSDMA_GLO_RX_BUSY		BIT(3)
+#define HSDMA_GLO_RX_DMA		BIT(2)
+#define HSDMA_GLO_TX_BUSY		BIT(1)
+#define HSDMA_GLO_TX_DMA		BIT(0)
+
+#define HSDMA_BT_SIZE_16BYTES		(0 << HSDMA_GLO_BT_SHIFT)
+#define HSDMA_BT_SIZE_32BYTES		(1 << HSDMA_GLO_BT_SHIFT)
+#define HSDMA_BT_SIZE_64BYTES		(2 << HSDMA_GLO_BT_SHIFT)
+#define HSDMA_BT_SIZE_128BYTES		(3 << HSDMA_GLO_BT_SHIFT)
+
+#define HSDMA_GLO_DEFAULT		(HSDMA_GLO_MULTI_DMA | \
+		HSDMA_GLO_RX_DMA | HSDMA_GLO_TX_DMA | HSDMA_BT_SIZE_32BYTES)
+
+/* HSDMA_REG_RST_CFG */
+#define HSDMA_RST_RX_SHIFT		16
+#define HSDMA_RST_TX_SHIFT		0
+
+/* HSDMA_REG_DELAY_INT */
+#define HSDMA_DELAY_INT_EN		BIT(15)
+#define HSDMA_DELAY_PEND_OFFSET		8
+#define HSDMA_DELAY_TIME_OFFSET		0
+#define HSDMA_DELAY_TX_OFFSET		16
+#define HSDMA_DELAY_RX_OFFSET		0
+
+#define HSDMA_DELAY_INIT(x)		(HSDMA_DELAY_INT_EN | \
+		((x) << HSDMA_DELAY_PEND_OFFSET))
+#define HSDMA_DELAY(x)			((HSDMA_DELAY_INIT(x) << \
+		HSDMA_DELAY_TX_OFFSET) | HSDMA_DELAY_INIT(x))
+
+/* HSDMA_REG_INT_STATUS */
+#define HSDMA_INT_DELAY_RX_COH		BIT(31)
+#define HSDMA_INT_DELAY_RX_INT		BIT(30)
+#define HSDMA_INT_DELAY_TX_COH		BIT(29)
+#define HSDMA_INT_DELAY_TX_INT		BIT(28)
+#define HSDMA_INT_RX_MASK		0x3
+#define HSDMA_INT_RX_SHIFT		16
+#define HSDMA_INT_RX_Q0			BIT(16)
+#define HSDMA_INT_TX_MASK		0xf
+#define HSDMA_INT_TX_SHIFT		0
+#define HSDMA_INT_TX_Q0			BIT(0)
+
+/* tx/rx dma desc flags */
+#define HSDMA_PLEN_MASK			0x3fff
+#define HSDMA_DESC_DONE			BIT(31)
+#define HSDMA_DESC_LS0			BIT(30)
+#define HSDMA_DESC_PLEN0(_x)		(((_x) & HSDMA_PLEN_MASK) << 16)
+#define HSDMA_DESC_TAG			BIT(15)
+#define HSDMA_DESC_LS1			BIT(14)
+#define HSDMA_DESC_PLEN1(_x)		((_x) & HSDMA_PLEN_MASK)
+
+/* align 4 bytes */
+#define HSDMA_ALIGN_SIZE		3
+/* align size 128bytes */
+#define HSDMA_MAX_PLEN			0x3f80
+
+struct hsdma_desc {
+	u32 addr0;
+	u32 flags;
+	u32 addr1;
+	u32 unused;
+};
+
+struct mtk_hsdma_sg {
+	dma_addr_t src_addr;
+	dma_addr_t dst_addr;
+	u32 len;
+};
+
+struct mtk_hsdma_desc {
+	struct virt_dma_desc vdesc;
+	unsigned int num_sgs;
+	struct mtk_hsdma_sg sg[1];
+};
+
+struct mtk_hsdma_chan {
+	struct virt_dma_chan vchan;
+	unsigned int id;
+	dma_addr_t desc_addr;
+	int tx_idx;
+	int rx_idx;
+	struct hsdma_desc *tx_ring;
+	struct hsdma_desc *rx_ring;
+	struct mtk_hsdma_desc *desc;
+	unsigned int next_sg;
+};
+
+struct mtk_hsdam_engine {
+	struct dma_device ddev;
+	struct device_dma_parameters dma_parms;
+	void __iomem *base;
+	struct tasklet_struct task;
+	volatile unsigned long chan_issued;
+
+	struct mtk_hsdma_chan chan[1];
+};
+
+static inline struct mtk_hsdam_engine *mtk_hsdma_chan_get_dev(
+		struct mtk_hsdma_chan *chan)
+{
+	return container_of(chan->vchan.chan.device, struct mtk_hsdam_engine,
+			ddev);
+}
+
+static inline struct mtk_hsdma_chan *to_mtk_hsdma_chan(struct dma_chan *c)
+{
+	return container_of(c, struct mtk_hsdma_chan, vchan.chan);
+}
+
+static inline struct mtk_hsdma_desc *to_mtk_hsdma_desc(
+		struct virt_dma_desc *vdesc)
+{
+	return container_of(vdesc, struct mtk_hsdma_desc, vdesc);
+}
+
+static inline u32 mtk_hsdma_read(struct mtk_hsdam_engine *hsdma, u32 reg)
+{
+	return readl(hsdma->base + reg);
+}
+
+static inline void mtk_hsdma_write(struct mtk_hsdam_engine *hsdma,
+		unsigned reg, u32 val)
+{
+	writel(val, hsdma->base + reg);
+}
+
+static void mtk_hsdma_reset_chan(struct mtk_hsdam_engine *hsdma,
+		struct mtk_hsdma_chan *chan)
+{
+	chan->tx_idx = 0;
+	chan->rx_idx = HSDMA_DESCS_NUM - 1;
+
+	mtk_hsdma_write(hsdma, HSDMA_REG_TX_CTX, chan->tx_idx);
+	mtk_hsdma_write(hsdma, HSDMA_REG_RX_CRX, chan->rx_idx);
+
+	mtk_hsdma_write(hsdma, HSDMA_REG_RST_CFG,
+			0x1 << (chan->id + HSDMA_RST_TX_SHIFT));
+	mtk_hsdma_write(hsdma, HSDMA_REG_RST_CFG,
+			0x1 << (chan->id + HSDMA_RST_RX_SHIFT));
+}
+
+static void hsdma_dump_reg(struct mtk_hsdam_engine *hsdma)
+{
+	dev_dbg(hsdma->ddev.dev, "tbase %08x, tcnt %08x, " \
+			"tctx %08x, tdtx: %08x, rbase %08x, " \
+			"rcnt %08x, rctx %08x, rdtx %08x\n",
+			mtk_hsdma_read(hsdma, HSDMA_REG_TX_BASE),
+			mtk_hsdma_read(hsdma, HSDMA_REG_TX_CNT),
+			mtk_hsdma_read(hsdma, HSDMA_REG_TX_CTX),
+			mtk_hsdma_read(hsdma, HSDMA_REG_TX_DTX),
+			mtk_hsdma_read(hsdma, HSDMA_REG_RX_BASE),
+			mtk_hsdma_read(hsdma, HSDMA_REG_RX_CNT),
+			mtk_hsdma_read(hsdma, HSDMA_REG_RX_CRX),
+			mtk_hsdma_read(hsdma, HSDMA_REG_RX_DRX));
+
+	dev_dbg(hsdma->ddev.dev, "info %08x, glo %08x, delay %08x, " \
+			"intr_stat %08x, intr_mask %08x\n",
+			mtk_hsdma_read(hsdma, HSDMA_REG_INFO),
+			mtk_hsdma_read(hsdma, HSDMA_REG_GLO_CFG),
+			mtk_hsdma_read(hsdma, HSDMA_REG_DELAY_INT),
+			mtk_hsdma_read(hsdma, HSDMA_REG_INT_STATUS),
+			mtk_hsdma_read(hsdma, HSDMA_REG_INT_MASK));
+}
+
+static void hsdma_dump_desc(struct mtk_hsdam_engine *hsdma,
+		struct mtk_hsdma_chan *chan)
+{
+	struct hsdma_desc *tx_desc;
+	struct hsdma_desc *rx_desc;
+	int i;
+
+	dev_dbg(hsdma->ddev.dev, "tx idx: %d, rx idx: %d\n",
+			chan->tx_idx, chan->rx_idx);
+
+	for (i = 0; i < HSDMA_DESCS_NUM; i++) {
+		tx_desc = &chan->tx_ring[i];
+		rx_desc = &chan->rx_ring[i];
+
+		dev_dbg(hsdma->ddev.dev, "%d tx addr0: %08x, flags %08x, " \
+				"tx addr1: %08x, rx addr0 %08x, flags %08x\n",
+				i, tx_desc->addr0, tx_desc->flags, \
+				tx_desc->addr1, rx_desc->addr0, rx_desc->flags);
+	}
+}
+
+static void mtk_hsdma_reset(struct mtk_hsdam_engine *hsdma,
+		struct mtk_hsdma_chan *chan)
+{
+	int i;
+
+	/* disable dma */
+	mtk_hsdma_write(hsdma, HSDMA_REG_GLO_CFG, 0);
+
+	/* disable intr */
+	mtk_hsdma_write(hsdma, HSDMA_REG_INT_MASK, 0);
+
+	/* init desc value */
+	for (i = 0; i < HSDMA_DESCS_NUM; i++) {
+		chan->tx_ring[i].addr0 = 0;
+		chan->tx_ring[i].flags = HSDMA_DESC_LS0 |
+			HSDMA_DESC_DONE;
+	}
+	for (i = 0; i < HSDMA_DESCS_NUM; i++) {
+		chan->rx_ring[i].addr0 = 0;
+		chan->rx_ring[i].flags = 0;
+	}
+
+	/* reset */
+	mtk_hsdma_reset_chan(hsdma, chan);
+
+	/* enable intr */
+	mtk_hsdma_write(hsdma, HSDMA_REG_INT_MASK, HSDMA_INT_RX_Q0);
+
+	/* enable dma */
+	mtk_hsdma_write(hsdma, HSDMA_REG_GLO_CFG, HSDMA_GLO_DEFAULT);
+}
+
+static int mtk_hsdma_terminate_all(struct dma_chan *c)
+{
+	struct mtk_hsdma_chan *chan = to_mtk_hsdma_chan(c);
+	struct mtk_hsdam_engine *hsdma = mtk_hsdma_chan_get_dev(chan);
+	unsigned long timeout;
+	LIST_HEAD(head);
+
+	spin_lock_bh(&chan->vchan.lock);
+	chan->desc = NULL;
+	clear_bit(chan->id, &hsdma->chan_issued);
+	vchan_get_all_descriptors(&chan->vchan, &head);
+	spin_unlock_bh(&chan->vchan.lock);
+
+	vchan_dma_desc_free_list(&chan->vchan, &head);
+
+	/* wait dma transfer complete */
+	timeout = jiffies + msecs_to_jiffies(2000);
+	while (mtk_hsdma_read(hsdma, HSDMA_REG_GLO_CFG) &
+			(HSDMA_GLO_RX_BUSY | HSDMA_GLO_TX_BUSY)) {
+		if (time_after_eq(jiffies, timeout)) {
+			hsdma_dump_desc(hsdma, chan);
+			mtk_hsdma_reset(hsdma, chan);
+			dev_err(hsdma->ddev.dev, "timeout, reset it\n");
+			break;
+		}
+		cpu_relax();
+	}
+
+	return 0;
+}
+
+static int mtk_hsdma_start_transfer(struct mtk_hsdam_engine *hsdma,
+		struct mtk_hsdma_chan *chan)
+{
+	dma_addr_t src, dst;
+	size_t len, tlen;
+	struct hsdma_desc *tx_desc, *rx_desc;
+	struct mtk_hsdma_sg *sg;
+	unsigned int i;
+	int rx_idx;
+
+	sg = &chan->desc->sg[0];
+	len = sg->len;
+	chan->desc->num_sgs = DIV_ROUND_UP(len, HSDMA_MAX_PLEN);
+
+	/* tx desc */
+	src = sg->src_addr;
+	for (i = 0; i < chan->desc->num_sgs; i++) {
+		if (len > HSDMA_MAX_PLEN)
+			tlen = HSDMA_MAX_PLEN;
+		else
+			tlen = len;
+
+		if (i & 0x1) {
+			tx_desc->addr1 = src;
+			tx_desc->flags |= HSDMA_DESC_PLEN1(tlen);
+		} else {
+			tx_desc = &chan->tx_ring[chan->tx_idx];
+			tx_desc->addr0 = src;
+			tx_desc->flags = HSDMA_DESC_PLEN0(tlen);
+
+			/* update index */
+			chan->tx_idx = HSDMA_NEXT_DESC(chan->tx_idx);
+		}
+
+		src += tlen;
+		len -= tlen;
+	}
+	if (i & 0x1)
+		tx_desc->flags |= HSDMA_DESC_LS0;
+	else
+		tx_desc->flags |= HSDMA_DESC_LS1;
+
+	/* rx desc */
+	rx_idx = HSDMA_NEXT_DESC(chan->rx_idx);
+	len = sg->len;
+	dst = sg->dst_addr;
+	for (i = 0; i < chan->desc->num_sgs; i++) {
+		rx_desc = &chan->rx_ring[rx_idx];
+		if (len > HSDMA_MAX_PLEN)
+			tlen = HSDMA_MAX_PLEN;
+		else
+			tlen = len;
+
+		rx_desc->addr0 = dst;
+		rx_desc->flags = HSDMA_DESC_PLEN0(tlen);
+
+		dst += tlen;
+		len -= tlen;
+
+		/* update index */
+		rx_idx = HSDMA_NEXT_DESC(rx_idx);
+	}
+
+	/* make sure desc and index all up to date */
+	wmb();
+	mtk_hsdma_write(hsdma, HSDMA_REG_TX_CTX, chan->tx_idx);
+
+	return 0;
+}
+
+static int gdma_next_desc(struct mtk_hsdma_chan *chan)
+{
+	struct virt_dma_desc *vdesc;
+
+	vdesc = vchan_next_desc(&chan->vchan);
+	if (!vdesc) {
+		chan->desc = NULL;
+		return 0;
+	}
+	chan->desc = to_mtk_hsdma_desc(vdesc);
+	chan->next_sg = 0;
+
+	return 1;
+}
+
+static void mtk_hsdma_chan_done(struct mtk_hsdam_engine *hsdma,
+		struct mtk_hsdma_chan *chan)
+{
+	struct mtk_hsdma_desc *desc;
+	int chan_issued;
+
+	chan_issued = 0;
+	spin_lock_bh(&chan->vchan.lock);
+	desc = chan->desc;
+	if (likely(desc)) {
+		if (chan->next_sg == desc->num_sgs) {
+			list_del(&desc->vdesc.node);
+			vchan_cookie_complete(&desc->vdesc);
+			chan_issued = gdma_next_desc(chan);
+		}
+	} else
+		dev_dbg(hsdma->ddev.dev, "no desc to complete\n");
+
+	if (chan_issued)
+		set_bit(chan->id, &hsdma->chan_issued);
+	spin_unlock_bh(&chan->vchan.lock);
+}
+
+static irqreturn_t mtk_hsdma_irq(int irq, void *devid)
+{
+	struct mtk_hsdam_engine *hsdma = devid;
+	u32 status;
+
+	status = mtk_hsdma_read(hsdma, HSDMA_REG_INT_STATUS);
+	if (unlikely(!status))
+		return IRQ_NONE;
+
+	if (likely(status & HSDMA_INT_RX_Q0))
+		tasklet_schedule(&hsdma->task);
+	else
+		dev_dbg(hsdma->ddev.dev, "unhandle irq status %08x\n",
+				status);
+	/* clean intr bits */
+	mtk_hsdma_write(hsdma, HSDMA_REG_INT_STATUS, status);
+
+	return IRQ_HANDLED;
+}
+
+static void mtk_hsdma_issue_pending(struct dma_chan *c)
+{
+	struct mtk_hsdma_chan *chan = to_mtk_hsdma_chan(c);
+	struct mtk_hsdam_engine *hsdma = mtk_hsdma_chan_get_dev(chan);
+
+	spin_lock_bh(&chan->vchan.lock);
+	if (vchan_issue_pending(&chan->vchan) && !chan->desc) {
+		if (gdma_next_desc(chan)) {
+			set_bit(chan->id, &hsdma->chan_issued);
+			tasklet_schedule(&hsdma->task);
+		} else
+			dev_dbg(hsdma->ddev.dev, "no desc to issue\n");
+	}
+	spin_unlock_bh(&chan->vchan.lock);
+}
+
+static struct dma_async_tx_descriptor * mtk_hsdma_prep_dma_memcpy(
+		struct dma_chan *c, dma_addr_t dest, dma_addr_t src,
+		size_t len, unsigned long flags)
+{
+	struct mtk_hsdma_chan *chan = to_mtk_hsdma_chan(c);
+	struct mtk_hsdma_desc *desc;
+
+	if (len <= 0)
+		return NULL;
+
+	desc = kzalloc(sizeof(struct mtk_hsdma_desc), GFP_ATOMIC);
+	if (!desc) {
+		dev_err(c->device->dev, "alloc memcpy decs error\n");
+		return NULL;
+	}
+
+	desc->sg[0].src_addr = src;
+	desc->sg[0].dst_addr = dest;
+	desc->sg[0].len = len;
+
+	return vchan_tx_prep(&chan->vchan, &desc->vdesc, flags);
+}
+
+static enum dma_status mtk_hsdma_tx_status(struct dma_chan *c,
+		dma_cookie_t cookie, struct dma_tx_state *state)
+{
+	return dma_cookie_status(c, cookie, state);
+}
+
+static void mtk_hsdma_free_chan_resources(struct dma_chan *c)
+{
+	vchan_free_chan_resources(to_virt_chan(c));
+}
+
+static void mtk_hsdma_desc_free(struct virt_dma_desc *vdesc)
+{
+	kfree(container_of(vdesc, struct mtk_hsdma_desc, vdesc));
+}
+
+static void mtk_hsdma_tx(struct mtk_hsdam_engine *hsdma)
+{
+	struct mtk_hsdma_chan *chan;
+
+	if (test_and_clear_bit(0, &hsdma->chan_issued)) {
+		chan = &hsdma->chan[0];
+		if (chan->desc) {
+			mtk_hsdma_start_transfer(hsdma, chan);
+		} else
+			dev_dbg(hsdma->ddev.dev,"chan 0 no desc to issue\n");
+	}
+}
+
+static void mtk_hsdma_rx(struct mtk_hsdam_engine *hsdma)
+{
+	struct mtk_hsdma_chan *chan;
+	int next_idx, drx_idx, cnt;
+
+	chan = &hsdma->chan[0];
+	next_idx = HSDMA_NEXT_DESC(chan->rx_idx);
+	drx_idx = mtk_hsdma_read(hsdma, HSDMA_REG_RX_DRX);
+
+	cnt = (drx_idx - next_idx) & HSDMA_DESCS_MASK;
+	if (!cnt)
+		return;
+
+	chan->next_sg += cnt;
+	chan->rx_idx = (chan->rx_idx + cnt) & HSDMA_DESCS_MASK;
+
+	/* update rx crx */
+	wmb();
+	mtk_hsdma_write(hsdma, HSDMA_REG_RX_CRX, chan->rx_idx);
+
+	mtk_hsdma_chan_done(hsdma, chan);
+}
+
+static void mtk_hsdma_tasklet(unsigned long arg)
+{
+	struct mtk_hsdam_engine *hsdma = (struct mtk_hsdam_engine *)arg;
+
+	mtk_hsdma_rx(hsdma);
+	mtk_hsdma_tx(hsdma);
+}
+
+static int mtk_hsdam_alloc_desc(struct mtk_hsdam_engine *hsdma,
+		struct mtk_hsdma_chan *chan)
+{
+	int i;
+
+	chan->tx_ring = dma_alloc_coherent(hsdma->ddev.dev,
+			2 * HSDMA_DESCS_NUM * sizeof(*chan->tx_ring),
+			&chan->desc_addr, GFP_ATOMIC | __GFP_ZERO);
+	if (!chan->tx_ring)
+		goto no_mem;
+
+	chan->rx_ring = &chan->tx_ring[HSDMA_DESCS_NUM];
+
+	/* init tx ring value */
+	for (i = 0; i < HSDMA_DESCS_NUM; i++)
+		chan->tx_ring[i].flags = HSDMA_DESC_LS0 | HSDMA_DESC_DONE;
+
+	return 0;
+no_mem:
+	return -ENOMEM;
+}
+
+static void mtk_hsdam_free_desc(struct mtk_hsdam_engine *hsdma,
+		struct mtk_hsdma_chan *chan)
+{
+	if (chan->tx_ring) {
+		dma_free_coherent(hsdma->ddev.dev,
+				2 * HSDMA_DESCS_NUM * sizeof(*chan->tx_ring),
+				chan->tx_ring, chan->desc_addr);
+		chan->tx_ring = NULL;
+		chan->rx_ring = NULL;
+	}
+}
+
+static int mtk_hsdma_init(struct mtk_hsdam_engine *hsdma)
+{
+	struct mtk_hsdma_chan *chan;
+	int ret;
+	u32 reg;
+
+	/* init desc */
+	chan = &hsdma->chan[0];
+	ret = mtk_hsdam_alloc_desc(hsdma, chan);
+	if (ret)
+		return ret;
+
+	/* tx */
+	mtk_hsdma_write(hsdma, HSDMA_REG_TX_BASE, chan->desc_addr);
+	mtk_hsdma_write(hsdma, HSDMA_REG_TX_CNT, HSDMA_DESCS_NUM);
+	/* rx */
+	mtk_hsdma_write(hsdma, HSDMA_REG_RX_BASE, chan->desc_addr +
+			(sizeof(struct hsdma_desc) * HSDMA_DESCS_NUM));
+	mtk_hsdma_write(hsdma, HSDMA_REG_RX_CNT, HSDMA_DESCS_NUM);
+	/* reset */
+	mtk_hsdma_reset_chan(hsdma, chan);
+
+	/* enable rx intr */
+	mtk_hsdma_write(hsdma, HSDMA_REG_INT_MASK, HSDMA_INT_RX_Q0);
+
+	/* enable dma */
+	mtk_hsdma_write(hsdma, HSDMA_REG_GLO_CFG, HSDMA_GLO_DEFAULT);
+
+	/* hardware info */
+	reg = mtk_hsdma_read(hsdma, HSDMA_REG_INFO);
+	dev_info(hsdma->ddev.dev, "rx: %d, tx: %d\n",
+			(reg >> HSDMA_INFO_RX_SHIFT) & HSDMA_INFO_RX_MASK,
+			(reg >> HSDMA_INFO_TX_SHIFT) & HSDMA_INFO_TX_MASK);
+
+	hsdma_dump_reg(hsdma);
+
+	return ret;
+}
+
+static void mtk_hsdma_uninit(struct mtk_hsdam_engine *hsdma)
+{
+	struct mtk_hsdma_chan *chan;
+
+	/* disable dma */
+	mtk_hsdma_write(hsdma, HSDMA_REG_GLO_CFG, 0);
+
+	/* disable intr */
+	mtk_hsdma_write(hsdma, HSDMA_REG_INT_MASK, 0);
+
+	/* free desc */
+	chan = &hsdma->chan[0];
+	mtk_hsdam_free_desc(hsdma, chan);
+
+	/* tx */
+	mtk_hsdma_write(hsdma, HSDMA_REG_TX_BASE, 0);
+	mtk_hsdma_write(hsdma, HSDMA_REG_TX_CNT, 0);
+	/* rx */
+	mtk_hsdma_write(hsdma, HSDMA_REG_RX_BASE, 0);
+	mtk_hsdma_write(hsdma, HSDMA_REG_RX_CNT, 0);
+	/* reset */
+	mtk_hsdma_reset_chan(hsdma, chan);
+}
+
+static const struct of_device_id mtk_hsdma_of_match[] = {
+	{ .compatible = "mediatek,mt7621-hsdma" },
+	{ },
+};
+
+static int mtk_hsdma_probe(struct platform_device *pdev)
+{
+	const struct of_device_id *match;
+	struct mtk_hsdma_chan *chan;
+	struct mtk_hsdam_engine *hsdma;
+	struct dma_device *dd;
+	struct resource *res;
+	int ret;
+	int irq;
+	void __iomem *base;
+
+	ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
+	if (ret)
+		return ret;
+
+	match = of_match_device(mtk_hsdma_of_match, &pdev->dev);
+	if (!match)
+		return -EINVAL;
+
+	hsdma = devm_kzalloc(&pdev->dev, sizeof(*hsdma), GFP_KERNEL);
+	if (!hsdma) {
+		dev_err(&pdev->dev, "alloc dma device failed\n");
+		return -EINVAL;
+	}
+
+	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+	base = devm_ioremap_resource(&pdev->dev, res);
+	if (IS_ERR(base))
+		return PTR_ERR(base);
+	hsdma->base = base + HSDMA_BASE_OFFSET;
+	tasklet_init(&hsdma->task, mtk_hsdma_tasklet, (unsigned long)hsdma);
+
+	irq = platform_get_irq(pdev, 0);
+	if (irq < 0) {
+		dev_err(&pdev->dev, "failed to get irq\n");
+		return -EINVAL;
+	}
+	ret = devm_request_irq(&pdev->dev, irq, mtk_hsdma_irq,
+			0, dev_name(&pdev->dev), hsdma);
+	if (ret) {
+		dev_err(&pdev->dev, "failed to request irq\n");
+		return ret;
+	}
+
+	device_reset(&pdev->dev);
+
+	dd = &hsdma->ddev;
+	dma_cap_set(DMA_MEMCPY, dd->cap_mask);
+	dd->copy_align = HSDMA_ALIGN_SIZE;
+	dd->device_free_chan_resources = mtk_hsdma_free_chan_resources;
+	dd->device_prep_dma_memcpy = mtk_hsdma_prep_dma_memcpy;
+	dd->device_terminate_all = mtk_hsdma_terminate_all;
+	dd->device_tx_status = mtk_hsdma_tx_status;
+	dd->device_issue_pending = mtk_hsdma_issue_pending;
+	dd->dev = &pdev->dev;
+	dd->dev->dma_parms = &hsdma->dma_parms;
+	dma_set_max_seg_size(dd->dev, HSDMA_MAX_PLEN);
+	INIT_LIST_HEAD(&dd->channels);
+
+	chan = &hsdma->chan[0];
+	chan->id = 0;
+	chan->vchan.desc_free = mtk_hsdma_desc_free;
+	vchan_init(&chan->vchan, dd);
+
+	/* init hardware */
+	ret = mtk_hsdma_init(hsdma);
+	if (ret) {
+		dev_err(&pdev->dev, "failed to alloc ring descs\n");
+		return ret;
+	}
+
+	ret = dma_async_device_register(dd);
+	if (ret) {
+		dev_err(&pdev->dev, "failed to register dma device\n");
+		return ret;
+	}
+
+	ret = of_dma_controller_register(pdev->dev.of_node,
+			of_dma_xlate_by_chan_id, hsdma);
+	if (ret) {
+		dev_err(&pdev->dev, "failed to register of dma controller\n");
+		goto err_unregister;
+	}
+
+	platform_set_drvdata(pdev, hsdma);
+
+	return 0;
+
+err_unregister:
+	dma_async_device_unregister(dd);
+	return ret;
+}
+
+static int mtk_hsdma_remove(struct platform_device *pdev)
+{
+	struct mtk_hsdam_engine *hsdma = platform_get_drvdata(pdev);
+
+	mtk_hsdma_uninit(hsdma);
+
+	of_dma_controller_free(pdev->dev.of_node);
+	dma_async_device_unregister(&hsdma->ddev);
+
+	return 0;
+}
+
+static struct platform_driver mtk_hsdma_driver = {
+	.probe = mtk_hsdma_probe,
+	.remove = mtk_hsdma_remove,
+	.driver = {
+		.name = "hsdma-mt7621",
+		.of_match_table = mtk_hsdma_of_match,
+	},
+};
+module_platform_driver(mtk_hsdma_driver);
+
+MODULE_AUTHOR("Michael Lee <igvtee@gmail.com>");
+MODULE_DESCRIPTION("MTK HSDMA driver");
+MODULE_LICENSE("GPL v2");
diff --git a/drivers/staging/mt7621-dma/ralink-gdma.c b/drivers/staging/mt7621-dma/ralink-gdma.c
new file mode 100644
index 0000000..23ddd64
--- /dev/null
+++ b/drivers/staging/mt7621-dma/ralink-gdma.c
@@ -0,0 +1,928 @@
+/*
+ *  Copyright (C) 2013, Lars-Peter Clausen <lars@metafoo.de>
+ *  GDMA4740 DMAC support
+ *
+ *  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/dmaengine.h>
+#include <linux/dma-mapping.h>
+#include <linux/err.h>
+#include <linux/init.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/irq.h>
+#include <linux/of_dma.h>
+#include <linux/reset.h>
+#include <linux/of_device.h>
+
+#include "virt-dma.h"
+
+#define GDMA_REG_SRC_ADDR(x)		(0x00 + (x) * 0x10)
+#define GDMA_REG_DST_ADDR(x)		(0x04 + (x) * 0x10)
+
+#define GDMA_REG_CTRL0(x)		(0x08 + (x) * 0x10)
+#define GDMA_REG_CTRL0_TX_MASK		0xffff
+#define GDMA_REG_CTRL0_TX_SHIFT		16
+#define GDMA_REG_CTRL0_CURR_MASK	0xff
+#define GDMA_REG_CTRL0_CURR_SHIFT	8
+#define	GDMA_REG_CTRL0_SRC_ADDR_FIXED	BIT(7)
+#define GDMA_REG_CTRL0_DST_ADDR_FIXED	BIT(6)
+#define GDMA_REG_CTRL0_BURST_MASK	0x7
+#define GDMA_REG_CTRL0_BURST_SHIFT	3
+#define	GDMA_REG_CTRL0_DONE_INT		BIT(2)
+#define	GDMA_REG_CTRL0_ENABLE		BIT(1)
+#define GDMA_REG_CTRL0_SW_MODE          BIT(0)
+
+#define GDMA_REG_CTRL1(x)		(0x0c + (x) * 0x10)
+#define GDMA_REG_CTRL1_SEG_MASK		0xf
+#define GDMA_REG_CTRL1_SEG_SHIFT	22
+#define GDMA_REG_CTRL1_REQ_MASK		0x3f
+#define GDMA_REG_CTRL1_SRC_REQ_SHIFT	16
+#define GDMA_REG_CTRL1_DST_REQ_SHIFT	8
+#define GDMA_REG_CTRL1_CONTINOUS	BIT(14)
+#define GDMA_REG_CTRL1_NEXT_MASK	0x1f
+#define GDMA_REG_CTRL1_NEXT_SHIFT	3
+#define GDMA_REG_CTRL1_COHERENT		BIT(2)
+#define GDMA_REG_CTRL1_FAIL		BIT(1)
+#define GDMA_REG_CTRL1_MASK		BIT(0)
+
+#define GDMA_REG_UNMASK_INT		0x200
+#define GDMA_REG_DONE_INT		0x204
+
+#define GDMA_REG_GCT			0x220
+#define GDMA_REG_GCT_CHAN_MASK		0x3
+#define GDMA_REG_GCT_CHAN_SHIFT		3
+#define GDMA_REG_GCT_VER_MASK		0x3
+#define GDMA_REG_GCT_VER_SHIFT		1
+#define GDMA_REG_GCT_ARBIT_RR		BIT(0)
+
+#define GDMA_REG_REQSTS			0x2a0
+#define GDMA_REG_ACKSTS			0x2a4
+#define GDMA_REG_FINSTS			0x2a8
+
+/* for RT305X gdma registers */
+#define GDMA_RT305X_CTRL0_REQ_MASK	0xf
+#define GDMA_RT305X_CTRL0_SRC_REQ_SHIFT	12
+#define GDMA_RT305X_CTRL0_DST_REQ_SHIFT	8
+
+#define GDMA_RT305X_CTRL1_FAIL		BIT(4)
+#define GDMA_RT305X_CTRL1_NEXT_MASK	0x7
+#define GDMA_RT305X_CTRL1_NEXT_SHIFT	1
+
+#define GDMA_RT305X_STATUS_INT		0x80
+#define GDMA_RT305X_STATUS_SIGNAL	0x84
+#define GDMA_RT305X_GCT			0x88
+
+/* for MT7621 gdma registers */
+#define GDMA_REG_PERF_START(x)		(0x230 + (x) * 0x8)
+#define GDMA_REG_PERF_END(x)		(0x234 + (x) * 0x8)
+
+enum gdma_dma_transfer_size {
+	GDMA_TRANSFER_SIZE_4BYTE	= 0,
+	GDMA_TRANSFER_SIZE_8BYTE	= 1,
+	GDMA_TRANSFER_SIZE_16BYTE	= 2,
+	GDMA_TRANSFER_SIZE_32BYTE	= 3,
+	GDMA_TRANSFER_SIZE_64BYTE	= 4,
+};
+
+struct gdma_dma_sg {
+	dma_addr_t src_addr;
+	dma_addr_t dst_addr;
+	u32 len;
+};
+
+struct gdma_dma_desc {
+	struct virt_dma_desc vdesc;
+
+	enum dma_transfer_direction direction;
+	bool cyclic;
+
+	u32 residue;
+	unsigned int num_sgs;
+	struct gdma_dma_sg sg[];
+};
+
+struct gdma_dmaengine_chan {
+	struct virt_dma_chan vchan;
+	unsigned int id;
+	unsigned int slave_id;
+
+	dma_addr_t fifo_addr;
+	enum gdma_dma_transfer_size burst_size;
+
+	struct gdma_dma_desc *desc;
+	unsigned int next_sg;
+};
+
+struct gdma_dma_dev {
+	struct dma_device ddev;
+	struct device_dma_parameters dma_parms;
+	struct gdma_data *data;
+	void __iomem *base;
+	struct tasklet_struct task;
+	volatile unsigned long chan_issued;
+	atomic_t cnt;
+
+	struct gdma_dmaengine_chan chan[];
+};
+
+struct gdma_data
+{
+	int chancnt;
+	u32 done_int_reg;
+	void (*init)(struct gdma_dma_dev *dma_dev);
+	int (*start_transfer)(struct gdma_dmaengine_chan *chan);
+};
+
+static struct gdma_dma_dev *gdma_dma_chan_get_dev(
+	struct gdma_dmaengine_chan *chan)
+{
+	return container_of(chan->vchan.chan.device, struct gdma_dma_dev,
+		ddev);
+}
+
+static struct gdma_dmaengine_chan *to_gdma_dma_chan(struct dma_chan *c)
+{
+	return container_of(c, struct gdma_dmaengine_chan, vchan.chan);
+}
+
+static struct gdma_dma_desc *to_gdma_dma_desc(struct virt_dma_desc *vdesc)
+{
+	return container_of(vdesc, struct gdma_dma_desc, vdesc);
+}
+
+static inline uint32_t gdma_dma_read(struct gdma_dma_dev *dma_dev,
+	unsigned int reg)
+{
+	return readl(dma_dev->base + reg);
+}
+
+static inline void gdma_dma_write(struct gdma_dma_dev *dma_dev,
+	unsigned reg, uint32_t val)
+{
+	writel(val, dma_dev->base + reg);
+}
+
+static struct gdma_dma_desc *gdma_dma_alloc_desc(unsigned int num_sgs)
+{
+	return kzalloc(sizeof(struct gdma_dma_desc) +
+		sizeof(struct gdma_dma_sg) * num_sgs, GFP_ATOMIC);
+}
+
+static enum gdma_dma_transfer_size gdma_dma_maxburst(u32 maxburst)
+{
+	if (maxburst < 2)
+		return GDMA_TRANSFER_SIZE_4BYTE;
+	else if (maxburst < 4)
+		return GDMA_TRANSFER_SIZE_8BYTE;
+	else if (maxburst < 8)
+		return GDMA_TRANSFER_SIZE_16BYTE;
+	else if (maxburst < 16)
+		return GDMA_TRANSFER_SIZE_32BYTE;
+	else
+		return GDMA_TRANSFER_SIZE_64BYTE;
+}
+
+static int gdma_dma_config(struct dma_chan *c,
+		struct dma_slave_config *config)
+{
+	struct gdma_dmaengine_chan *chan = to_gdma_dma_chan(c);
+	struct gdma_dma_dev *dma_dev = gdma_dma_chan_get_dev(chan);
+
+	if (config->device_fc) {
+		dev_err(dma_dev->ddev.dev, "not support flow controller\n");
+		return -EINVAL;
+	}
+
+	switch (config->direction) {
+	case DMA_MEM_TO_DEV:
+		if (config->dst_addr_width != DMA_SLAVE_BUSWIDTH_4_BYTES) {
+			dev_err(dma_dev->ddev.dev, "only support 4 byte buswidth\n");
+			return -EINVAL;
+		}
+		chan->slave_id = config->slave_id;
+		chan->fifo_addr = config->dst_addr;
+		chan->burst_size = gdma_dma_maxburst(config->dst_maxburst);
+		break;
+	case DMA_DEV_TO_MEM:
+		if (config->src_addr_width != DMA_SLAVE_BUSWIDTH_4_BYTES) {
+			dev_err(dma_dev->ddev.dev, "only support 4 byte buswidth\n");
+			return -EINVAL;
+		}
+		chan->slave_id = config->slave_id;
+		chan->fifo_addr = config->src_addr;
+		chan->burst_size = gdma_dma_maxburst(config->src_maxburst);
+		break;
+	default:
+		dev_err(dma_dev->ddev.dev, "direction type %d error\n",
+				config->direction);
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+static int gdma_dma_terminate_all(struct dma_chan *c)
+{
+	struct gdma_dmaengine_chan *chan = to_gdma_dma_chan(c);
+	struct gdma_dma_dev *dma_dev = gdma_dma_chan_get_dev(chan);
+	unsigned long flags, timeout;
+	LIST_HEAD(head);
+	int i = 0;
+
+	spin_lock_irqsave(&chan->vchan.lock, flags);
+	chan->desc = NULL;
+	clear_bit(chan->id, &dma_dev->chan_issued);
+	vchan_get_all_descriptors(&chan->vchan, &head);
+	spin_unlock_irqrestore(&chan->vchan.lock, flags);
+
+	vchan_dma_desc_free_list(&chan->vchan, &head);
+
+	/* wait dma transfer complete */
+	timeout = jiffies + msecs_to_jiffies(5000);
+	while (gdma_dma_read(dma_dev, GDMA_REG_CTRL0(chan->id)) &
+			GDMA_REG_CTRL0_ENABLE) {
+		if (time_after_eq(jiffies, timeout)) {
+			dev_err(dma_dev->ddev.dev, "chan %d wait timeout\n",
+					chan->id);
+			/* restore to init value */
+			gdma_dma_write(dma_dev, GDMA_REG_CTRL0(chan->id), 0);
+			break;
+		}
+		cpu_relax();
+		i++;
+	}
+
+	if (i)
+		dev_dbg(dma_dev->ddev.dev, "terminate chan %d loops %d\n",
+				chan->id, i);
+
+	return 0;
+}
+
+static void rt305x_dump_reg(struct gdma_dma_dev *dma_dev, int id)
+{
+	dev_dbg(dma_dev->ddev.dev, "chan %d, src %08x, dst %08x, ctr0 %08x, " \
+			"ctr1 %08x, intr %08x, signal %08x\n", id,
+			gdma_dma_read(dma_dev, GDMA_REG_SRC_ADDR(id)),
+			gdma_dma_read(dma_dev, GDMA_REG_DST_ADDR(id)),
+			gdma_dma_read(dma_dev, GDMA_REG_CTRL0(id)),
+			gdma_dma_read(dma_dev, GDMA_REG_CTRL1(id)),
+			gdma_dma_read(dma_dev, GDMA_RT305X_STATUS_INT),
+			gdma_dma_read(dma_dev, GDMA_RT305X_STATUS_SIGNAL));
+}
+
+static int rt305x_gdma_start_transfer(struct gdma_dmaengine_chan *chan)
+{
+	struct gdma_dma_dev *dma_dev = gdma_dma_chan_get_dev(chan);
+	dma_addr_t src_addr, dst_addr;
+	struct gdma_dma_sg *sg;
+	uint32_t ctrl0, ctrl1;
+
+	/* verify chan is already stopped */
+	ctrl0 = gdma_dma_read(dma_dev, GDMA_REG_CTRL0(chan->id));
+	if (unlikely(ctrl0 & GDMA_REG_CTRL0_ENABLE)) {
+		dev_err(dma_dev->ddev.dev, "chan %d is start(%08x).\n",
+				chan->id, ctrl0);
+		rt305x_dump_reg(dma_dev, chan->id);
+		return -EINVAL;
+	}
+
+	sg = &chan->desc->sg[chan->next_sg];
+	if (chan->desc->direction == DMA_MEM_TO_DEV) {
+		src_addr = sg->src_addr;
+		dst_addr = chan->fifo_addr;
+		ctrl0 = GDMA_REG_CTRL0_DST_ADDR_FIXED | \
+			(8 << GDMA_RT305X_CTRL0_SRC_REQ_SHIFT) | \
+			(chan->slave_id << GDMA_RT305X_CTRL0_DST_REQ_SHIFT);
+	} else if (chan->desc->direction == DMA_DEV_TO_MEM) {
+		src_addr = chan->fifo_addr;
+		dst_addr = sg->dst_addr;
+		ctrl0 = GDMA_REG_CTRL0_SRC_ADDR_FIXED | \
+			(chan->slave_id << GDMA_RT305X_CTRL0_SRC_REQ_SHIFT) | \
+			(8 << GDMA_RT305X_CTRL0_DST_REQ_SHIFT);
+	} else if (chan->desc->direction == DMA_MEM_TO_MEM) {
+		/*
+		 * TODO: memcpy function have bugs. sometime it will copy
+		 * more 8 bytes data when using dmatest verify.
+		 */
+		src_addr = sg->src_addr;
+		dst_addr = sg->dst_addr;
+		ctrl0 = GDMA_REG_CTRL0_SW_MODE | \
+			(8 << GDMA_REG_CTRL1_SRC_REQ_SHIFT) | \
+			(8 << GDMA_REG_CTRL1_DST_REQ_SHIFT);
+	} else {
+		dev_err(dma_dev->ddev.dev, "direction type %d error\n",
+				chan->desc->direction);
+		return -EINVAL;
+	}
+
+	ctrl0 |= (sg->len << GDMA_REG_CTRL0_TX_SHIFT) | \
+		 (chan->burst_size << GDMA_REG_CTRL0_BURST_SHIFT) | \
+		 GDMA_REG_CTRL0_DONE_INT | GDMA_REG_CTRL0_ENABLE;
+	ctrl1 = chan->id << GDMA_REG_CTRL1_NEXT_SHIFT;
+
+	chan->next_sg++;
+	gdma_dma_write(dma_dev, GDMA_REG_SRC_ADDR(chan->id), src_addr);
+	gdma_dma_write(dma_dev, GDMA_REG_DST_ADDR(chan->id), dst_addr);
+	gdma_dma_write(dma_dev, GDMA_REG_CTRL1(chan->id), ctrl1);
+
+	/* make sure next_sg is update */
+	wmb();
+	gdma_dma_write(dma_dev, GDMA_REG_CTRL0(chan->id), ctrl0);
+
+	return 0;
+}
+
+static void rt3883_dump_reg(struct gdma_dma_dev *dma_dev, int id)
+{
+	dev_dbg(dma_dev->ddev.dev, "chan %d, src %08x, dst %08x, ctr0 %08x, " \
+			"ctr1 %08x, unmask %08x, done %08x, " \
+			"req %08x, ack %08x, fin %08x\n", id,
+			gdma_dma_read(dma_dev, GDMA_REG_SRC_ADDR(id)),
+			gdma_dma_read(dma_dev, GDMA_REG_DST_ADDR(id)),
+			gdma_dma_read(dma_dev, GDMA_REG_CTRL0(id)),
+			gdma_dma_read(dma_dev, GDMA_REG_CTRL1(id)),
+			gdma_dma_read(dma_dev, GDMA_REG_UNMASK_INT),
+			gdma_dma_read(dma_dev, GDMA_REG_DONE_INT),
+			gdma_dma_read(dma_dev, GDMA_REG_REQSTS),
+			gdma_dma_read(dma_dev, GDMA_REG_ACKSTS),
+			gdma_dma_read(dma_dev, GDMA_REG_FINSTS));
+}
+
+static int rt3883_gdma_start_transfer(struct gdma_dmaengine_chan *chan)
+{
+	struct gdma_dma_dev *dma_dev = gdma_dma_chan_get_dev(chan);
+	dma_addr_t src_addr, dst_addr;
+	struct gdma_dma_sg *sg;
+	uint32_t ctrl0, ctrl1;
+
+	/* verify chan is already stopped */
+	ctrl0 = gdma_dma_read(dma_dev, GDMA_REG_CTRL0(chan->id));
+	if (unlikely(ctrl0 & GDMA_REG_CTRL0_ENABLE)) {
+		dev_err(dma_dev->ddev.dev, "chan %d is start(%08x).\n",
+				chan->id, ctrl0);
+		rt3883_dump_reg(dma_dev, chan->id);
+		return -EINVAL;
+	}
+
+	sg = &chan->desc->sg[chan->next_sg];
+	if (chan->desc->direction == DMA_MEM_TO_DEV) {
+		src_addr = sg->src_addr;
+		dst_addr = chan->fifo_addr;
+		ctrl0 = GDMA_REG_CTRL0_DST_ADDR_FIXED;
+		ctrl1 = (32 << GDMA_REG_CTRL1_SRC_REQ_SHIFT) | \
+			(chan->slave_id << GDMA_REG_CTRL1_DST_REQ_SHIFT);
+	} else if (chan->desc->direction == DMA_DEV_TO_MEM) {
+		src_addr = chan->fifo_addr;
+		dst_addr = sg->dst_addr;
+		ctrl0 = GDMA_REG_CTRL0_SRC_ADDR_FIXED;
+		ctrl1 = (chan->slave_id << GDMA_REG_CTRL1_SRC_REQ_SHIFT) | \
+			(32 << GDMA_REG_CTRL1_DST_REQ_SHIFT) | \
+			GDMA_REG_CTRL1_COHERENT;
+	} else if (chan->desc->direction == DMA_MEM_TO_MEM) {
+		src_addr = sg->src_addr;
+		dst_addr = sg->dst_addr;
+		ctrl0 = GDMA_REG_CTRL0_SW_MODE;
+		ctrl1 = (32 << GDMA_REG_CTRL1_SRC_REQ_SHIFT) | \
+			(32 << GDMA_REG_CTRL1_DST_REQ_SHIFT) | \
+			GDMA_REG_CTRL1_COHERENT;
+	} else {
+		dev_err(dma_dev->ddev.dev, "direction type %d error\n",
+				chan->desc->direction);
+		return -EINVAL;
+	}
+
+	ctrl0 |= (sg->len << GDMA_REG_CTRL0_TX_SHIFT) | \
+		 (chan->burst_size << GDMA_REG_CTRL0_BURST_SHIFT) | \
+		 GDMA_REG_CTRL0_DONE_INT | GDMA_REG_CTRL0_ENABLE;
+	ctrl1 |= chan->id << GDMA_REG_CTRL1_NEXT_SHIFT;
+
+	chan->next_sg++;
+	gdma_dma_write(dma_dev, GDMA_REG_SRC_ADDR(chan->id), src_addr);
+	gdma_dma_write(dma_dev, GDMA_REG_DST_ADDR(chan->id), dst_addr);
+	gdma_dma_write(dma_dev, GDMA_REG_CTRL1(chan->id), ctrl1);
+
+	/* make sure next_sg is update */
+	wmb();
+	gdma_dma_write(dma_dev, GDMA_REG_CTRL0(chan->id), ctrl0);
+
+	return 0;
+}
+
+static inline int gdma_start_transfer(struct gdma_dma_dev *dma_dev,
+		struct gdma_dmaengine_chan *chan)
+{
+	return dma_dev->data->start_transfer(chan);
+}
+
+static int gdma_next_desc(struct gdma_dmaengine_chan *chan)
+{
+	struct virt_dma_desc *vdesc;
+
+	vdesc = vchan_next_desc(&chan->vchan);
+	if (!vdesc) {
+		chan->desc = NULL;
+		return 0;
+	}
+	chan->desc = to_gdma_dma_desc(vdesc);
+	chan->next_sg = 0;
+
+	return 1;
+}
+
+static void gdma_dma_chan_irq(struct gdma_dma_dev *dma_dev,
+		struct gdma_dmaengine_chan *chan)
+{
+	struct gdma_dma_desc *desc;
+	unsigned long flags;
+	int chan_issued;
+
+	chan_issued = 0;
+	spin_lock_irqsave(&chan->vchan.lock, flags);
+	desc = chan->desc;
+	if (desc) {
+		if (desc->cyclic) {
+			vchan_cyclic_callback(&desc->vdesc);
+			if (chan->next_sg == desc->num_sgs)
+				chan->next_sg = 0;
+			chan_issued = 1;
+		} else {
+			desc->residue -= desc->sg[chan->next_sg - 1].len;
+			if (chan->next_sg == desc->num_sgs) {
+				list_del(&desc->vdesc.node);
+				vchan_cookie_complete(&desc->vdesc);
+				chan_issued = gdma_next_desc(chan);
+			} else
+				chan_issued = 1;
+		}
+	} else
+		dev_dbg(dma_dev->ddev.dev, "chan %d no desc to complete\n",
+				chan->id);
+	if (chan_issued)
+		set_bit(chan->id, &dma_dev->chan_issued);
+	spin_unlock_irqrestore(&chan->vchan.lock, flags);
+}
+
+static irqreturn_t gdma_dma_irq(int irq, void *devid)
+{
+	struct gdma_dma_dev *dma_dev = devid;
+	u32 done, done_reg;
+	unsigned int i;
+
+	done_reg = dma_dev->data->done_int_reg;
+	done = gdma_dma_read(dma_dev, done_reg);
+	if (unlikely(!done))
+		return IRQ_NONE;
+
+	/* clean done bits */
+	gdma_dma_write(dma_dev, done_reg, done);
+
+	i = 0;
+	while (done) {
+		if (done & 0x1) {
+			gdma_dma_chan_irq(dma_dev, &dma_dev->chan[i]);
+			atomic_dec(&dma_dev->cnt);
+		}
+		done >>= 1;
+		i++;
+	}
+
+	/* start only have work to do */
+	if (dma_dev->chan_issued)
+		tasklet_schedule(&dma_dev->task);
+
+	return IRQ_HANDLED;
+}
+
+static void gdma_dma_issue_pending(struct dma_chan *c)
+{
+	struct gdma_dmaengine_chan *chan = to_gdma_dma_chan(c);
+	struct gdma_dma_dev *dma_dev = gdma_dma_chan_get_dev(chan);
+	unsigned long flags;
+
+	spin_lock_irqsave(&chan->vchan.lock, flags);
+	if (vchan_issue_pending(&chan->vchan) && !chan->desc) {
+		if (gdma_next_desc(chan)) {
+			set_bit(chan->id, &dma_dev->chan_issued);
+			tasklet_schedule(&dma_dev->task);
+		} else
+			dev_dbg(dma_dev->ddev.dev, "chan %d no desc to issue\n",
+					chan->id);
+	}
+	spin_unlock_irqrestore(&chan->vchan.lock, flags);
+}
+
+static struct dma_async_tx_descriptor *gdma_dma_prep_slave_sg(
+		struct dma_chan *c, struct scatterlist *sgl,
+		unsigned int sg_len, enum dma_transfer_direction direction,
+		unsigned long flags, void *context)
+{
+	struct gdma_dmaengine_chan *chan = to_gdma_dma_chan(c);
+	struct gdma_dma_desc *desc;
+	struct scatterlist *sg;
+	unsigned int i;
+
+	desc = gdma_dma_alloc_desc(sg_len);
+	if (!desc) {
+		dev_err(c->device->dev, "alloc sg decs error\n");
+		return NULL;
+	}
+	desc->residue = 0;
+
+	for_each_sg(sgl, sg, sg_len, i) {
+		if (direction == DMA_MEM_TO_DEV)
+			desc->sg[i].src_addr = sg_dma_address(sg);
+		else if (direction == DMA_DEV_TO_MEM)
+			desc->sg[i].dst_addr = sg_dma_address(sg);
+		else {
+			dev_err(c->device->dev, "direction type %d error\n",
+					direction);
+			goto free_desc;
+		}
+
+		if (unlikely(sg_dma_len(sg) > GDMA_REG_CTRL0_TX_MASK)) {
+			dev_err(c->device->dev, "sg len too large %d\n",
+					sg_dma_len(sg));
+			goto free_desc;
+		}
+		desc->sg[i].len = sg_dma_len(sg);
+		desc->residue += sg_dma_len(sg);
+	}
+
+	desc->num_sgs = sg_len;
+	desc->direction = direction;
+	desc->cyclic = false;
+
+	return vchan_tx_prep(&chan->vchan, &desc->vdesc, flags);
+
+free_desc:
+	kfree(desc);
+	return NULL;
+}
+
+static struct dma_async_tx_descriptor * gdma_dma_prep_dma_memcpy(
+		struct dma_chan *c, dma_addr_t dest, dma_addr_t src,
+		size_t len, unsigned long flags)
+{
+	struct gdma_dmaengine_chan *chan = to_gdma_dma_chan(c);
+	struct gdma_dma_desc *desc;
+	unsigned int num_periods, i;
+	size_t xfer_count;
+
+	if (len <= 0)
+		return NULL;
+
+	chan->burst_size = gdma_dma_maxburst(len >> 2);
+
+	xfer_count = GDMA_REG_CTRL0_TX_MASK;
+	num_periods = DIV_ROUND_UP(len, xfer_count);
+
+	desc = gdma_dma_alloc_desc(num_periods);
+	if (!desc) {
+		dev_err(c->device->dev, "alloc memcpy decs error\n");
+		return NULL;
+	}
+	desc->residue = len;
+
+	for (i = 0; i < num_periods; i++) {
+		desc->sg[i].src_addr = src;
+		desc->sg[i].dst_addr = dest;
+		if (len > xfer_count) {
+			desc->sg[i].len = xfer_count;
+		} else {
+			desc->sg[i].len = len;
+		}
+		src += desc->sg[i].len;
+		dest += desc->sg[i].len;
+		len -= desc->sg[i].len;
+	}
+
+	desc->num_sgs = num_periods;
+	desc->direction = DMA_MEM_TO_MEM;
+	desc->cyclic = false;
+
+	return vchan_tx_prep(&chan->vchan, &desc->vdesc, flags);
+}
+
+static struct dma_async_tx_descriptor *gdma_dma_prep_dma_cyclic(
+	struct dma_chan *c, dma_addr_t buf_addr, size_t buf_len,
+	size_t period_len, enum dma_transfer_direction direction,
+	unsigned long flags)
+{
+	struct gdma_dmaengine_chan *chan = to_gdma_dma_chan(c);
+	struct gdma_dma_desc *desc;
+	unsigned int num_periods, i;
+
+	if (buf_len % period_len)
+		return NULL;
+
+	if (period_len > GDMA_REG_CTRL0_TX_MASK) {
+		dev_err(c->device->dev, "cyclic len too large %d\n",
+				period_len);
+		return NULL;
+	}
+
+	num_periods = buf_len / period_len;
+	desc = gdma_dma_alloc_desc(num_periods);
+	if (!desc) {
+		dev_err(c->device->dev, "alloc cyclic decs error\n");
+		return NULL;
+	}
+	desc->residue = buf_len;
+
+	for (i = 0; i < num_periods; i++) {
+		if (direction == DMA_MEM_TO_DEV)
+			desc->sg[i].src_addr = buf_addr;
+		else if (direction == DMA_DEV_TO_MEM)
+			desc->sg[i].dst_addr = buf_addr;
+		else {
+			dev_err(c->device->dev, "direction type %d error\n",
+					direction);
+			goto free_desc;
+		}
+		desc->sg[i].len = period_len;
+		buf_addr += period_len;
+	}
+
+	desc->num_sgs = num_periods;
+	desc->direction = direction;
+	desc->cyclic = true;
+
+	return vchan_tx_prep(&chan->vchan, &desc->vdesc, flags);
+
+free_desc:
+	kfree(desc);
+	return NULL;
+}
+
+static enum dma_status gdma_dma_tx_status(struct dma_chan *c,
+	dma_cookie_t cookie, struct dma_tx_state *state)
+{
+	struct gdma_dmaengine_chan *chan = to_gdma_dma_chan(c);
+	struct virt_dma_desc *vdesc;
+	enum dma_status status;
+	unsigned long flags;
+	struct gdma_dma_desc *desc;
+
+	status = dma_cookie_status(c, cookie, state);
+	if (status == DMA_COMPLETE || !state)
+		return status;
+
+	spin_lock_irqsave(&chan->vchan.lock, flags);
+	desc = chan->desc;
+	if (desc && (cookie == desc->vdesc.tx.cookie)) {
+		/*
+		 * We never update edesc->residue in the cyclic case, so we
+		 * can tell the remaining room to the end of the circular
+		 * buffer.
+		 */
+		if (desc->cyclic)
+			state->residue = desc->residue -
+				((chan->next_sg - 1) * desc->sg[0].len);
+		else
+			state->residue = desc->residue;
+	} else if ((vdesc = vchan_find_desc(&chan->vchan, cookie)))
+		state->residue = to_gdma_dma_desc(vdesc)->residue;
+	spin_unlock_irqrestore(&chan->vchan.lock, flags);
+
+	dev_dbg(c->device->dev, "tx residue %d bytes\n", state->residue);
+
+	return status;
+}
+
+static void gdma_dma_free_chan_resources(struct dma_chan *c)
+{
+	vchan_free_chan_resources(to_virt_chan(c));
+}
+
+static void gdma_dma_desc_free(struct virt_dma_desc *vdesc)
+{
+	kfree(container_of(vdesc, struct gdma_dma_desc, vdesc));
+}
+
+static void gdma_dma_tasklet(unsigned long arg)
+{
+	struct gdma_dma_dev *dma_dev = (struct gdma_dma_dev *)arg;
+	struct gdma_dmaengine_chan *chan;
+	static unsigned int last_chan;
+	unsigned int i, chan_mask;
+
+	/* record last chan to round robin all chans */
+	i = last_chan;
+	chan_mask = dma_dev->data->chancnt - 1;
+	do {
+		/*
+		 * on mt7621. when verify with dmatest with all
+		 * channel is enable. we need to limit only two
+		 * channel is working at the same time. otherwise the
+		 * data will have problem.
+		 */
+		if (atomic_read(&dma_dev->cnt) >= 2) {
+			last_chan = i;
+			break;
+		}
+
+		if (test_and_clear_bit(i, &dma_dev->chan_issued)) {
+			chan = &dma_dev->chan[i];
+			if (chan->desc) {
+				atomic_inc(&dma_dev->cnt);
+				gdma_start_transfer(dma_dev, chan);
+			} else
+				dev_dbg(dma_dev->ddev.dev, "chan %d no desc to issue\n", chan->id);
+
+			if (!dma_dev->chan_issued)
+				break;
+		}
+
+		i = (i + 1) & chan_mask;
+	} while (i != last_chan);
+}
+
+static void rt305x_gdma_init(struct gdma_dma_dev *dma_dev)
+{
+	uint32_t gct;
+
+	/* all chans round robin */
+	gdma_dma_write(dma_dev, GDMA_RT305X_GCT, GDMA_REG_GCT_ARBIT_RR);
+
+	gct = gdma_dma_read(dma_dev, GDMA_RT305X_GCT);
+	dev_info(dma_dev->ddev.dev, "revision: %d, channels: %d\n",
+			(gct >> GDMA_REG_GCT_VER_SHIFT) & GDMA_REG_GCT_VER_MASK,
+			8 << ((gct >> GDMA_REG_GCT_CHAN_SHIFT) &
+				GDMA_REG_GCT_CHAN_MASK));
+}
+
+static void rt3883_gdma_init(struct gdma_dma_dev *dma_dev)
+{
+	uint32_t gct;
+
+	/* all chans round robin */
+	gdma_dma_write(dma_dev, GDMA_REG_GCT, GDMA_REG_GCT_ARBIT_RR);
+
+	gct = gdma_dma_read(dma_dev, GDMA_REG_GCT);
+	dev_info(dma_dev->ddev.dev, "revision: %d, channels: %d\n",
+			(gct >> GDMA_REG_GCT_VER_SHIFT) & GDMA_REG_GCT_VER_MASK,
+			8 << ((gct >> GDMA_REG_GCT_CHAN_SHIFT) &
+				GDMA_REG_GCT_CHAN_MASK));
+}
+
+static struct gdma_data rt305x_gdma_data = {
+	.chancnt = 8,
+	.done_int_reg = GDMA_RT305X_STATUS_INT,
+	.init = rt305x_gdma_init,
+	.start_transfer = rt305x_gdma_start_transfer,
+};
+
+static struct gdma_data rt3883_gdma_data = {
+	.chancnt = 16,
+	.done_int_reg = GDMA_REG_DONE_INT,
+	.init = rt3883_gdma_init,
+	.start_transfer = rt3883_gdma_start_transfer,
+};
+
+static const struct of_device_id gdma_of_match_table[] = {
+	{ .compatible = "ralink,rt305x-gdma", .data = &rt305x_gdma_data },
+	{ .compatible = "ralink,rt3883-gdma", .data = &rt3883_gdma_data },
+	{ },
+};
+
+static int gdma_dma_probe(struct platform_device *pdev)
+{
+	const struct of_device_id *match;
+	struct gdma_dmaengine_chan *chan;
+	struct gdma_dma_dev *dma_dev;
+	struct dma_device *dd;
+	unsigned int i;
+	struct resource *res;
+	int ret;
+	int irq;
+	void __iomem *base;
+	struct gdma_data *data;
+
+	ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
+	if (ret)
+		return ret;
+
+	match = of_match_device(gdma_of_match_table, &pdev->dev);
+	if (!match)
+		return -EINVAL;
+	data = (struct gdma_data *) match->data;
+
+	dma_dev = devm_kzalloc(&pdev->dev, sizeof(*dma_dev) +
+			(sizeof(struct gdma_dmaengine_chan) * data->chancnt),
+			GFP_KERNEL);
+	if (!dma_dev) {
+		dev_err(&pdev->dev, "alloc dma device failed\n");
+		return -EINVAL;
+	}
+	dma_dev->data = data;
+
+	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+	base = devm_ioremap_resource(&pdev->dev, res);
+	if (IS_ERR(base))
+		return PTR_ERR(base);
+	dma_dev->base = base;
+	tasklet_init(&dma_dev->task, gdma_dma_tasklet, (unsigned long)dma_dev);
+
+	irq = platform_get_irq(pdev, 0);
+	if (irq < 0) {
+		dev_err(&pdev->dev, "failed to get irq\n");
+		return -EINVAL;
+	}
+	ret = devm_request_irq(&pdev->dev, irq, gdma_dma_irq,
+			0, dev_name(&pdev->dev), dma_dev);
+	if (ret) {
+		dev_err(&pdev->dev, "failed to request irq\n");
+		return ret;
+	}
+
+	device_reset(&pdev->dev);
+
+	dd = &dma_dev->ddev;
+	dma_cap_set(DMA_MEMCPY, dd->cap_mask);
+	dma_cap_set(DMA_SLAVE, dd->cap_mask);
+	dma_cap_set(DMA_CYCLIC, dd->cap_mask);
+	dd->device_free_chan_resources = gdma_dma_free_chan_resources;
+	dd->device_prep_dma_memcpy = gdma_dma_prep_dma_memcpy;
+	dd->device_prep_slave_sg = gdma_dma_prep_slave_sg;
+	dd->device_prep_dma_cyclic = gdma_dma_prep_dma_cyclic;
+	dd->device_config = gdma_dma_config;
+	dd->device_terminate_all = gdma_dma_terminate_all;
+	dd->device_tx_status = gdma_dma_tx_status;
+	dd->device_issue_pending = gdma_dma_issue_pending;
+
+	dd->src_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
+	dd->dst_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
+	dd->directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
+	dd->residue_granularity = DMA_RESIDUE_GRANULARITY_SEGMENT;
+
+	dd->dev = &pdev->dev;
+	dd->dev->dma_parms = &dma_dev->dma_parms;
+	dma_set_max_seg_size(dd->dev, GDMA_REG_CTRL0_TX_MASK);
+	INIT_LIST_HEAD(&dd->channels);
+
+	for (i = 0; i < data->chancnt; i++) {
+		chan = &dma_dev->chan[i];
+		chan->id = i;
+		chan->vchan.desc_free = gdma_dma_desc_free;
+		vchan_init(&chan->vchan, dd);
+	}
+
+	/* init hardware */
+	data->init(dma_dev);
+
+	ret = dma_async_device_register(dd);
+	if (ret) {
+		dev_err(&pdev->dev, "failed to register dma device\n");
+		return ret;
+	}
+
+	ret = of_dma_controller_register(pdev->dev.of_node,
+		of_dma_xlate_by_chan_id, dma_dev);
+	if (ret) {
+		dev_err(&pdev->dev, "failed to register of dma controller\n");
+		goto err_unregister;
+	}
+
+	platform_set_drvdata(pdev, dma_dev);
+
+	return 0;
+
+err_unregister:
+	dma_async_device_unregister(dd);
+	return ret;
+}
+
+static int gdma_dma_remove(struct platform_device *pdev)
+{
+	struct gdma_dma_dev *dma_dev = platform_get_drvdata(pdev);
+
+	tasklet_kill(&dma_dev->task);
+        of_dma_controller_free(pdev->dev.of_node);
+	dma_async_device_unregister(&dma_dev->ddev);
+
+	return 0;
+}
+
+static struct platform_driver gdma_dma_driver = {
+	.probe = gdma_dma_probe,
+	.remove = gdma_dma_remove,
+	.driver = {
+		.name = "gdma-rt2880",
+		.of_match_table = gdma_of_match_table,
+	},
+};
+module_platform_driver(gdma_dma_driver);
+
+MODULE_AUTHOR("Lars-Peter Clausen <lars@metafoo.de>");
+MODULE_DESCRIPTION("Ralink/MTK DMA driver");
+MODULE_LICENSE("GPL v2");