From nobody Mon Sep 17 00:00:00 2001
From: H�vard Skinnemoen <hskinnemoen@atmel.com>
Date: Fri Dec 2 13:24:24 2005 +0100
Subject: [PATCH] AVR32: DesignWare DMA Controller

This patch adds a driver for the Synopsys DesignWare DMA Controller.

---

 arch/avr32/Kconfig           |    4 
 arch/avr32/Makefile          |    1 
 arch/avr32/drivers/Makefile  |    1 
 arch/avr32/drivers/dw-dmac.c |  754 +++++++++++++++++++++++++++++++++++++++++++
 arch/avr32/drivers/dw-dmac.h |   42 ++
 5 files changed, 802 insertions(+)

Index: linux-2.6.18-avr32/arch/avr32/Makefile
===================================================================
--- linux-2.6.18-avr32.orig/arch/avr32/Makefile	2006-11-02 14:17:29.000000000 +0100
+++ linux-2.6.18-avr32/arch/avr32/Makefile	2006-11-02 15:53:13.000000000 +0100
@@ -30,6 +30,7 @@ core-$(CONFIG_BOARD_ATSTK1000)		+= arch/
 core-$(CONFIG_LOADER_U_BOOT)		+= arch/avr32/boot/u-boot/
 core-y					+= arch/avr32/kernel/
 core-y					+= arch/avr32/mm/
+drivers-y				+= arch/avr32/drivers/
 drivers-$(CONFIG_OPROFILE)		+= arch/avr32/oprofile/
 libs-y					+= arch/avr32/lib/
 
Index: linux-2.6.18-avr32/arch/avr32/drivers/Makefile
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ linux-2.6.18-avr32/arch/avr32/drivers/Makefile	2006-11-02 14:17:29.000000000 +0100
@@ -0,0 +1 @@
+obj-$(CONFIG_DW_DMAC)			+= dw-dmac.o
Index: linux-2.6.18-avr32/arch/avr32/drivers/dw-dmac.c
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ linux-2.6.18-avr32/arch/avr32/drivers/dw-dmac.c	2006-11-02 15:55:35.000000000 +0100
@@ -0,0 +1,754 @@
+/*
+ * Driver for the Synopsys DesignWare DMA Controller
+ *
+ * Copyright (C) 2005-2006 Atmel Corporation
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ */
+
+#include <linux/clk.h>
+#include <linux/device.h>
+#include <linux/dma-mapping.h>
+#include <linux/dmapool.h>
+#include <linux/init.h>
+#include <linux/interrupt.h>
+#include <linux/module.h>
+#include <linux/platform_device.h>
+
+#include <asm/dma-controller.h>
+#include <asm/io.h>
+
+#include "dw-dmac.h"
+
+#define DMAC_NR_CHANNELS 3
+#define DMAC_MAX_BLOCKSIZE 4095
+
+enum {
+	CH_STATE_FREE = 0,
+	CH_STATE_ALLOCATED,
+	CH_STATE_BUSY,
+};
+
+struct dw_dma_lli {
+	dma_addr_t	sar;
+	dma_addr_t	dar;
+	dma_addr_t	llp;
+	u32		ctllo;
+	u32		ctlhi;
+	u32		sstat;
+	u32		dstat;
+};
+
+struct dw_dma_block {
+	struct dw_dma_lli *lli_vaddr;
+	dma_addr_t lli_dma_addr;
+};
+
+struct dw_dma_channel {
+	unsigned int state;
+        int is_cyclic;
+	struct dma_request_sg *req_sg;
+	struct dma_request_cyclic *req_cyclic;
+	unsigned int nr_blocks;
+	int direction;
+	struct dw_dma_block *block;
+};
+
+struct dw_dma_controller {
+	spinlock_t lock;
+	void * __iomem	regs;
+	struct dma_pool *lli_pool;
+	struct clk *hclk;
+	struct dma_controller dma;
+	struct dw_dma_channel channel[DMAC_NR_CHANNELS];
+};
+#define to_dw_dmac(dmac) container_of(dmac, struct dw_dma_controller, dma)
+
+#define dmac_writel_hi(dmac, reg, value) \
+	__raw_writel((value), (dmac)->regs + DW_DMAC_##reg + 4)
+#define dmac_readl_hi(dmac, reg) \
+	__raw_readl((dmac)->regs + DW_DMAC_##reg + 4)
+#define dmac_writel_lo(dmac, reg, value) \
+	__raw_writel((value), (dmac)->regs + DW_DMAC_##reg)
+#define dmac_readl_lo(dmac, reg) \
+	__raw_readl((dmac)->regs + DW_DMAC_##reg)
+#define dmac_chan_writel_hi(dmac, chan, reg, value) \
+	__raw_writel((value), ((dmac)->regs + 0x58 * (chan) \
+			       + DW_DMAC_CHAN_##reg + 4))
+#define dmac_chan_readl_hi(dmac, chan, reg) \
+	__raw_readl((dmac)->regs + 0x58 * (chan) + DW_DMAC_CHAN_##reg + 4)
+#define dmac_chan_writel_lo(dmac, chan, reg, value) \
+	__raw_writel((value), (dmac)->regs + 0x58 * (chan) + DW_DMAC_CHAN_##reg)
+#define dmac_chan_readl_lo(dmac, chan, reg) \
+	__raw_readl((dmac)->regs + 0x58 * (chan) + DW_DMAC_CHAN_##reg)
+#define set_channel_bit(dmac, reg, chan) \
+	dmac_writel_lo(dmac, reg, (1 << (chan)) | (1 << ((chan) + 8)))
+#define clear_channel_bit(dmac, reg, chan) \
+	dmac_writel_lo(dmac, reg, (0 << (chan)) | (1 << ((chan) + 8)))
+
+static int dmac_alloc_channel(struct dma_controller *_dmac)
+{
+	struct dw_dma_controller *dmac = to_dw_dmac(_dmac);
+	struct dw_dma_channel *chan;
+	unsigned long flags;
+	int i;
+
+	spin_lock_irqsave(&dmac->lock, flags);
+	for (i = 0; i < DMAC_NR_CHANNELS; i++)
+		if (dmac->channel[i].state == CH_STATE_FREE)
+			break;
+
+	if (i < DMAC_NR_CHANNELS) {
+		chan = &dmac->channel[i];
+		chan->state = CH_STATE_ALLOCATED;
+	} else {
+		i = -EBUSY;
+	}
+
+	spin_unlock_irqrestore(&dmac->lock, flags);
+
+	return i;
+}
+
+static void dmac_release_channel(struct dma_controller *_dmac, int channel)
+{
+	struct dw_dma_controller *dmac = to_dw_dmac(_dmac);
+
+	BUG_ON(channel >= DMAC_NR_CHANNELS
+	       || dmac->channel[channel].state != CH_STATE_ALLOCATED);
+
+	dmac->channel[channel].state = CH_STATE_FREE;
+}
+
+static struct dw_dma_block *allocate_blocks(struct dw_dma_controller *dmac,
+					    unsigned int nr_blocks)
+{
+	struct dw_dma_block *block;
+	void *p;
+	unsigned int i;
+
+	block = kmalloc(nr_blocks * sizeof(*block),
+			GFP_KERNEL);
+	if (unlikely(!block))
+		return NULL;
+
+	for (i = 0; i < nr_blocks; i++) {
+		p = dma_pool_alloc(dmac->lli_pool, GFP_KERNEL,
+				   &block[i].lli_dma_addr);
+		block[i].lli_vaddr = p;
+		if (unlikely(!p))
+			goto fail;
+	}
+
+	return block;
+
+fail:
+	for (i = 0; i < nr_blocks; i++) {
+		if (!block[i].lli_vaddr)
+			break;
+		dma_pool_free(dmac->lli_pool, block[i].lli_vaddr,
+			      block[i].lli_dma_addr);
+	}
+	kfree(block);
+	return NULL;
+}
+
+static int dmac_prepare_request_sg(struct dma_controller *_dmac,
+				   struct dma_request_sg *req)
+{
+	struct dw_dma_controller *dmac = to_dw_dmac(_dmac);
+	struct dw_dma_channel *chan;
+	unsigned long ctlhi, ctllo, cfghi, cfglo;
+	unsigned long block_size;
+	int ret, i, nr_blocks, direction;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dmac->lock, flags);
+
+	ret = -EINVAL;
+	if (req->req.channel >= DMAC_NR_CHANNELS
+	    || dmac->channel[req->req.channel].state != CH_STATE_ALLOCATED
+	    || req->block_size > DMAC_MAX_BLOCKSIZE) {
+		spin_unlock_irqrestore(&dmac->lock, flags);
+		return -EINVAL;
+	}
+
+	chan = &dmac->channel[req->req.channel];
+	chan->state = CH_STATE_BUSY;
+	chan->req_sg = req;
+	chan->is_cyclic = 0;
+
+	/*
+	 * We have marked the channel as busy, so no need to keep the
+	 * lock as long as we only touch the channel-specific
+	 * registers
+	 */
+	spin_unlock_irqrestore(&dmac->lock, flags);
+
+	/*
+	 * There may be limitations in the driver and/or the DMA
+	 * controller that prevents us from sending a whole
+	 * scatterlist item in one go.  Taking this into account,
+	 * calculate the number of block transfers we need to set up.
+	 *
+	 * FIXME: Let the peripheral driver know about the maximum
+	 * block size we support. We really don't want to use a
+	 * different block size than what was suggested by the
+	 * peripheral.
+	 *
+	 * Each block will get its own Linked List Item (LLI) below.
+	 */
+	block_size = req->block_size;
+	pr_debug("block_size = %lu, nr_sg = %u\n", block_size, req->nr_sg);
+	for (i = 0, nr_blocks = 0; i < req->nr_sg; i++) {
+		pr_debug("sg[i].length = %u\n", req->sg[i].length);
+		BUG_ON(req->sg[i].length % block_size);
+		nr_blocks += req->sg[i].length / block_size;
+	}
+
+	BUG_ON(nr_blocks == 0);
+	chan->nr_blocks = nr_blocks;
+
+	ret = -EINVAL;
+	cfglo = cfghi = 0;
+	switch (req->direction) {
+	case DMA_DIR_MEM_TO_PERIPH:
+		direction = DMA_TO_DEVICE;
+		cfghi = req->periph_id << (43 - 32);
+		break;
+
+	case DMA_DIR_PERIPH_TO_MEM:
+		direction = DMA_FROM_DEVICE;
+		cfghi = req->periph_id << (39 - 32);
+		break;
+	default:
+		goto out_unclaim_channel;
+	}
+
+        chan->direction = direction;
+
+	dmac_chan_writel_hi(dmac, req->req.channel, CFG, cfghi);
+	dmac_chan_writel_lo(dmac, req->req.channel, CFG, cfglo);
+
+	ctlhi = block_size >> req->width;
+	ctllo = ((req->direction << 20)
+		 // | (1 << 14) | (1 << 11) // source/dest burst trans len
+		 | (req->width << 4) | (req->width << 1)
+		 | (1 << 0));		 // interrupt enable
+
+	if (nr_blocks == 1) {
+		/* Only one block: No need to use block chaining */
+		if (direction == DMA_TO_DEVICE) {
+			dmac_chan_writel_lo(dmac, req->req.channel, SAR,
+					    req->sg->dma_address);
+			dmac_chan_writel_lo(dmac, req->req.channel, DAR,
+					    req->data_reg);
+			ctllo |= 2 << 7; // no dst increment
+		} else {
+			dmac_chan_writel_lo(dmac, req->req.channel, SAR,
+					    req->data_reg);
+			dmac_chan_writel_lo(dmac, req->req.channel, DAR,
+					    req->sg->dma_address);
+			ctllo |= 2 << 9; // no src increment
+		}
+		dmac_chan_writel_lo(dmac, req->req.channel, CTL, ctllo);
+		dmac_chan_writel_hi(dmac, req->req.channel, CTL, ctlhi);
+	} else {
+		struct dw_dma_lli *lli, *lli_prev = NULL;
+		int j = 0, offset = 0;
+
+		ret = -ENOMEM;
+		chan->block = allocate_blocks(dmac, nr_blocks);
+		if (!chan->block)
+			goto out_unclaim_channel;
+
+		if (direction == DMA_TO_DEVICE)
+			ctllo |= 1 << 28 | 1 << 27 | 2 << 7;
+		else
+			ctllo |= 1 << 28 | 1 << 27 | 2 << 9;
+
+		/*
+		 * Map scatterlist items to blocks. One scatterlist
+		 * item may need more than one block for the reasons
+		 * mentioned above.
+		 */
+		for (i = 0; i < nr_blocks; i++) {
+			lli = chan->block[i].lli_vaddr;
+			if (lli_prev) {
+				lli_prev->llp = chan->block[i].lli_dma_addr;
+				pr_debug("lli[%d] (0x%p/0x%x): 0x%x 0x%x 0x%x 0x%x 0x%x\n",
+					 i - 1, chan->block[i - 1].lli_vaddr,
+					 chan->block[i - 1].lli_dma_addr,
+					 lli_prev->sar, lli_prev->dar, lli_prev->llp,
+					 lli_prev->ctllo, lli_prev->ctlhi);
+			}
+			lli->llp = 0;
+			lli->ctllo = ctllo;
+			lli->ctlhi = ctlhi;
+			if (direction == DMA_TO_DEVICE) {
+				lli->sar = req->sg[j].dma_address + offset;
+				lli->dar = req->data_reg;
+			} else {
+				lli->sar = req->data_reg;
+				lli->dar = req->sg[j].dma_address + offset;
+			}
+			lli_prev = lli;
+
+			offset += block_size;
+			if (offset > req->sg[j].length) {
+				j++;
+				offset = 0;
+			}
+		}
+
+		pr_debug("lli[%d] (0x%p/0x%x): 0x%x 0x%x 0x%x 0x%x 0x%x\n",
+			 i - 1, chan->block[i - 1].lli_vaddr,
+			 chan->block[i - 1].lli_dma_addr, lli_prev->sar,
+			 lli_prev->dar, lli_prev->llp,
+			 lli_prev->ctllo, lli_prev->ctlhi);
+
+		/*
+		 * SAR, DAR and CTL are initialized from the LLI. We
+		 * only have to enable the LLI bits in CTL.
+		 */
+		dmac_chan_writel_lo(dmac, req->req.channel, LLP,
+				    chan->block[0].lli_dma_addr);
+		dmac_chan_writel_lo(dmac, req->req.channel, CTL, 1 << 28 | 1 << 27);
+	}
+
+	set_channel_bit(dmac, MASK_XFER, req->req.channel);
+	set_channel_bit(dmac, MASK_ERROR, req->req.channel);
+	if (req->req.block_complete)
+		set_channel_bit(dmac, MASK_BLOCK, req->req.channel);
+	else
+		clear_channel_bit(dmac, MASK_BLOCK, req->req.channel);
+
+	return 0;
+
+out_unclaim_channel:
+	chan->state = CH_STATE_ALLOCATED;
+	return ret;
+}
+
+static int dmac_prepare_request_cyclic(struct dma_controller *_dmac,
+                                       struct dma_request_cyclic *req)
+{
+	struct dw_dma_controller *dmac = to_dw_dmac(_dmac);
+	struct dw_dma_channel *chan;
+	unsigned long ctlhi, ctllo, cfghi, cfglo;
+	unsigned long block_size;
+	int ret, i, direction;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dmac->lock, flags);
+
+        block_size = (req->buffer_size/req->periods) >> req->width;
+
+	ret = -EINVAL;
+	if (req->req.channel >= DMAC_NR_CHANNELS
+	    || dmac->channel[req->req.channel].state != CH_STATE_ALLOCATED
+            || (req->periods == 0)
+	    || block_size > DMAC_MAX_BLOCKSIZE) {
+		spin_unlock_irqrestore(&dmac->lock, flags);
+		return -EINVAL;
+	}
+
+	chan = &dmac->channel[req->req.channel];
+	chan->state = CH_STATE_BUSY;
+	chan->is_cyclic = 1;
+        chan->req_cyclic = req;
+
+	/*
+	 * We have marked the channel as busy, so no need to keep the
+	 * lock as long as we only touch the channel-specific
+	 * registers
+	 */
+	spin_unlock_irqrestore(&dmac->lock, flags);
+
+	/*
+          Setup
+	 */
+	BUG_ON(req->buffer_size % req->periods);
+	/* printk(KERN_INFO "block_size = %lu, periods = %u\n", block_size, req->periods); */
+
+	chan->nr_blocks = req->periods;
+
+	ret = -EINVAL;
+	cfglo = cfghi = 0;
+	switch (req->direction) {
+	case DMA_DIR_MEM_TO_PERIPH:
+		direction = DMA_TO_DEVICE;
+		cfghi = req->periph_id << (43 - 32);
+		break;
+
+	case DMA_DIR_PERIPH_TO_MEM:
+		direction = DMA_FROM_DEVICE;
+		cfghi = req->periph_id << (39 - 32);
+		break;
+	default:
+		goto out_unclaim_channel;
+	}
+
+        chan->direction = direction;
+
+	dmac_chan_writel_hi(dmac, req->req.channel, CFG, cfghi);
+	dmac_chan_writel_lo(dmac, req->req.channel, CFG, cfglo);
+
+	ctlhi = block_size;
+	ctllo = ((req->direction << 20)
+		 | (req->width << 4) | (req->width << 1)
+		 | (1 << 0));		 // interrupt enable
+
+        {
+		struct dw_dma_lli *lli = NULL, *lli_prev = NULL;
+
+		ret = -ENOMEM;
+		chan->block = allocate_blocks(dmac, req->periods);
+		if (!chan->block)
+			goto out_unclaim_channel;
+
+		if (direction == DMA_TO_DEVICE)
+			ctllo |= 1 << 28 | 1 << 27 | 2 << 7;
+		else
+			ctllo |= 1 << 28 | 1 << 27 | 2 << 9;
+
+		/*
+		 * Set up a linked list items where each period gets
+		 * an item. The linked list item for the last period
+		 * points back to the star of the buffer making a
+		 * cyclic buffer.
+		 */
+		for (i = 0; i < req->periods; i++) {
+			lli = chan->block[i].lli_vaddr;
+			if (lli_prev) {
+				lli_prev->llp = chan->block[i].lli_dma_addr;
+				/* printk(KERN_INFO "lli[%d] (0x%p/0x%x): 0x%x 0x%x 0x%x 0x%x 0x%x\n",
+				   i - 1, chan->block[i - 1].lli_vaddr,
+				   chan->block[i - 1].lli_dma_addr,
+				   lli_prev->sar, lli_prev->dar, lli_prev->llp,
+				   lli_prev->ctllo, lli_prev->ctlhi);*/
+			}
+			lli->llp = 0;
+			lli->ctllo = ctllo;
+			lli->ctlhi = ctlhi;
+			if (direction == DMA_TO_DEVICE) {
+				lli->sar = req->buffer_start + i*(block_size << req->width);
+				lli->dar = req->data_reg;
+			} else {
+				lli->sar = req->data_reg;
+				lli->dar = req->buffer_start + i*(block_size << req->width);
+			}
+			lli_prev = lli;
+		}
+		lli->llp = chan->block[0].lli_dma_addr;
+
+		/*printk(KERN_INFO "lli[%d] (0x%p/0x%x): 0x%x 0x%x 0x%x 0x%x 0x%x\n",
+		  i - 1, chan->block[i - 1].lli_vaddr,
+		  chan->block[i - 1].lli_dma_addr, lli_prev->sar,
+		  lli_prev->dar, lli_prev->llp,
+		  lli_prev->ctllo, lli_prev->ctlhi); */
+
+		/*
+		 * SAR, DAR and CTL are initialized from the LLI. We
+		 * only have to enable the LLI bits in CTL.
+		 */
+		dmac_chan_writel_lo(dmac, req->req.channel, LLP,
+				    chan->block[0].lli_dma_addr);
+		dmac_chan_writel_lo(dmac, req->req.channel, CTL, 1 << 28 | 1 << 27);
+	}
+
+	clear_channel_bit(dmac, MASK_XFER, req->req.channel);
+	set_channel_bit(dmac, MASK_ERROR, req->req.channel);
+	if (req->req.block_complete)
+		set_channel_bit(dmac, MASK_BLOCK, req->req.channel);
+	else
+		clear_channel_bit(dmac, MASK_BLOCK, req->req.channel);
+
+	return 0;
+
+out_unclaim_channel:
+	chan->state = CH_STATE_ALLOCATED;
+	return ret;
+}
+
+static int dmac_start_request(struct dma_controller *_dmac,
+			      unsigned int channel)
+{
+	struct dw_dma_controller *dmac = to_dw_dmac(_dmac);
+
+	BUG_ON(channel >= DMAC_NR_CHANNELS);
+
+	set_channel_bit(dmac, CH_EN, channel);
+
+	return 0;
+}
+
+static dma_addr_t dmac_get_current_pos(struct dma_controller *_dmac,
+                                       unsigned int channel)
+{
+	struct dw_dma_controller *dmac = to_dw_dmac(_dmac);
+	struct dw_dma_channel *chan;
+        dma_addr_t current_pos;
+
+	BUG_ON(channel >= DMAC_NR_CHANNELS);
+
+        chan = &dmac->channel[channel];
+
+	switch (chan->direction) {
+	case DMA_TO_DEVICE:
+		current_pos = dmac_chan_readl_lo(dmac, channel, SAR);
+		break;
+	case DMA_FROM_DEVICE:
+		current_pos = dmac_chan_readl_lo(dmac, channel, DAR);
+		break;
+	default:
+		return 0;
+	}
+
+
+        if (!current_pos) {
+		if (chan->is_cyclic) {
+			current_pos = chan->req_cyclic->buffer_start;
+		} else {
+			current_pos = chan->req_sg->sg->dma_address;
+		}
+	}
+
+	return current_pos;
+}
+
+
+static void cleanup_channel(struct dw_dma_controller *dmac,
+			    struct dw_dma_channel *chan)
+{
+	unsigned int i;
+
+	if (chan->nr_blocks > 1) {
+		for (i = 0; i < chan->nr_blocks; i++)
+			dma_pool_free(dmac->lli_pool, chan->block[i].lli_vaddr,
+				      chan->block[i].lli_dma_addr);
+		kfree(chan->block);
+	}
+
+	chan->state = CH_STATE_ALLOCATED;
+}
+
+static int dmac_stop_request(struct dma_controller *_dmac,
+                             unsigned int channel)
+{
+	struct dw_dma_controller *dmac = to_dw_dmac(_dmac);
+
+	BUG_ON(channel >= DMAC_NR_CHANNELS);
+
+	BUG_ON(dmac->channel[channel].state != CH_STATE_BUSY);
+
+	clear_channel_bit(dmac, CH_EN, channel);
+
+        cleanup_channel(dmac, &dmac->channel[channel]);
+
+	return 0;
+}
+
+
+static void dmac_block_complete(struct dw_dma_controller *dmac)
+{
+	struct dw_dma_channel *chan;
+	unsigned long status, chanid;
+
+	status = dmac_readl_lo(dmac, STATUS_BLOCK);
+
+	while (status) {
+		struct dma_request *req;
+		chanid = __ffs(status);
+		chan = &dmac->channel[chanid];
+
+                if (chan->is_cyclic) {
+			BUG_ON(!chan->req_cyclic
+			       || !chan->req_cyclic->req.block_complete);
+			req = &chan->req_cyclic->req;
+                } else {
+			BUG_ON(!chan->req_sg || !chan->req_sg->req.block_complete);
+			req = &chan->req_sg->req;
+                }
+		dmac_writel_lo(dmac, CLEAR_BLOCK, 1 << chanid);
+		req->block_complete(req);
+		status = dmac_readl_lo(dmac, STATUS_BLOCK);
+	}
+}
+
+static void dmac_xfer_complete(struct dw_dma_controller *dmac)
+{
+	struct dw_dma_channel *chan;
+	struct dma_request *req;
+	unsigned long status, chanid;
+
+	status = dmac_readl_lo(dmac, STATUS_XFER);
+
+	while (status) {
+		chanid = __ffs(status);
+		chan = &dmac->channel[chanid];
+
+		dmac_writel_lo(dmac, CLEAR_XFER, 1 << chanid);
+
+                req = &chan->req_sg->req;
+                BUG_ON(!req);
+                cleanup_channel(dmac, chan);
+                if (req->xfer_complete)
+			req->xfer_complete(req);
+
+		status = dmac_readl_lo(dmac, STATUS_XFER);
+	}
+}
+
+static void dmac_error(struct dw_dma_controller *dmac)
+{
+	struct dw_dma_channel *chan;
+	unsigned long status, chanid;
+
+	status = dmac_readl_lo(dmac, STATUS_ERROR);
+
+	while (status) {
+		struct dma_request *req;
+
+		chanid = __ffs(status);
+		chan = &dmac->channel[chanid];
+
+		dmac_writel_lo(dmac, CLEAR_ERROR, 1 << chanid);
+		clear_channel_bit(dmac, CH_EN, chanid);
+
+                if (chan->is_cyclic) {
+			BUG_ON(!chan->req_cyclic);
+			req = &chan->req_cyclic->req;
+                } else {
+			BUG_ON(!chan->req_sg);
+			req = &chan->req_sg->req;
+                }
+
+		cleanup_channel(dmac, chan);
+		if (req->error)
+			req->error(req);
+
+		status = dmac_readl_lo(dmac, STATUS_XFER);
+	}
+}
+
+static irqreturn_t dmac_interrupt(int irq, void *dev_id, struct pt_regs *regs)
+{
+	struct dw_dma_controller *dmac = dev_id;
+	unsigned long status;
+	int ret = IRQ_NONE;
+
+	spin_lock(&dmac->lock);
+
+	status = dmac_readl_lo(dmac, STATUS_INT);
+
+	while (status) {
+		ret = IRQ_HANDLED;
+		if (status & 0x10)
+			dmac_error(dmac);
+		if (status & 0x02)
+			dmac_block_complete(dmac);
+		if (status & 0x01)
+			dmac_xfer_complete(dmac);
+
+		status = dmac_readl_lo(dmac, STATUS_INT);
+	}
+
+	spin_unlock(&dmac->lock);
+	return ret;
+}
+
+static int __devinit dmac_probe(struct platform_device *pdev)
+{
+	struct dw_dma_controller *dmac;
+	struct resource *regs;
+	int ret;
+
+	regs = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+	if (!regs)
+		return -ENXIO;
+
+	dmac = kmalloc(sizeof(*dmac), GFP_KERNEL);
+	if (!dmac)
+		return -ENOMEM;
+	memset(dmac, 0, sizeof(*dmac));
+
+	dmac->hclk = clk_get(&pdev->dev, "hclk");
+	if (IS_ERR(dmac->hclk)) {
+		ret = PTR_ERR(dmac->hclk);
+		goto out_free_dmac;
+	}
+	clk_enable(dmac->hclk);
+
+	ret = -ENOMEM;
+	dmac->lli_pool = dma_pool_create("dmac", &pdev->dev,
+					 sizeof(struct dw_dma_lli), 4, 0);
+	if (!dmac->lli_pool)
+		goto out_disable_clk;
+
+	spin_lock_init(&dmac->lock);
+	dmac->dma.dev = &pdev->dev;
+	dmac->dma.alloc_channel = dmac_alloc_channel;
+	dmac->dma.release_channel = dmac_release_channel;
+	dmac->dma.prepare_request_sg = dmac_prepare_request_sg;
+	dmac->dma.prepare_request_cyclic = dmac_prepare_request_cyclic;
+	dmac->dma.start_request = dmac_start_request;
+	dmac->dma.stop_request = dmac_stop_request;
+	dmac->dma.get_current_pos = dmac_get_current_pos;
+
+	dmac->regs = ioremap(regs->start, regs->end - regs->start + 1);
+	if (!dmac->regs)
+		goto out_free_pool;
+
+	ret = request_irq(platform_get_irq(pdev, 0), dmac_interrupt,
+			  SA_SAMPLE_RANDOM, pdev->name, dmac);
+	if (ret)
+		goto out_unmap_regs;
+
+	/* Enable the DMA controller */
+	dmac_writel_lo(dmac, CFG, 1);
+
+	register_dma_controller(&dmac->dma);
+
+	printk(KERN_INFO
+	       "dmac%d: DesignWare DMA controller at 0x%p irq %d\n",
+	       dmac->dma.id, dmac->regs, platform_get_irq(pdev, 0));
+
+	return 0;
+
+out_unmap_regs:
+	iounmap(dmac->regs);
+out_free_pool:
+	dma_pool_destroy(dmac->lli_pool);
+out_disable_clk:
+	clk_disable(dmac->hclk);
+	clk_put(dmac->hclk);
+out_free_dmac:
+	kfree(dmac);
+	return ret;
+}
+
+static struct platform_driver dmac_driver = {
+	.probe		= dmac_probe,
+	.driver		= {
+		.name		= "dmac",
+	},
+};
+
+static int __init dmac_init(void)
+{
+	return platform_driver_register(&dmac_driver);
+}
+subsys_initcall(dmac_init);
+
+static void __exit dmac_exit(void)
+{
+	platform_driver_unregister(&dmac_driver);
+}
+module_exit(dmac_exit);
+
+MODULE_DESCRIPTION("Synopsys DesignWare DMA Controller driver");
+MODULE_AUTHOR("Haavard Skinnemoen <hskinnemoen@atmel.com>");
+MODULE_LICENSE("GPL");
Index: linux-2.6.18-avr32/arch/avr32/drivers/dw-dmac.h
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ linux-2.6.18-avr32/arch/avr32/drivers/dw-dmac.h	2006-11-02 14:17:29.000000000 +0100
@@ -0,0 +1,42 @@
+/*
+ * Driver for the Synopsys DesignWare DMA Controller
+ *
+ * Copyright (C) 2005-2006 Atmel Corporation
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ */
+#ifndef __AVR32_DW_DMAC_H__
+#define __AVR32_DW_DMAC_H__
+
+#define DW_DMAC_CFG		0x398
+#define DW_DMAC_CH_EN		0x3a0
+
+#define DW_DMAC_STATUS_XFER	0x2e8
+#define DW_DMAC_STATUS_BLOCK	0x2f0
+#define DW_DMAC_STATUS_ERROR	0x308
+
+#define DW_DMAC_MASK_XFER	0x310
+#define DW_DMAC_MASK_BLOCK	0x318
+#define DW_DMAC_MASK_ERROR	0x330
+
+#define DW_DMAC_CLEAR_XFER	0x338
+#define DW_DMAC_CLEAR_BLOCK	0x340
+#define DW_DMAC_CLEAR_ERROR	0x358
+
+#define DW_DMAC_STATUS_INT	0x360
+
+#define DW_DMAC_CHAN_SAR	0x000
+#define DW_DMAC_CHAN_DAR	0x008
+#define DW_DMAC_CHAN_LLP	0x010
+#define DW_DMAC_CHAN_CTL	0x018
+#define DW_DMAC_CHAN_SSTAT	0x020
+#define DW_DMAC_CHAN_DSTAT	0x028
+#define DW_DMAC_CHAN_SSTATAR	0x030
+#define DW_DMAC_CHAN_DSTATAR	0x038
+#define DW_DMAC_CHAN_CFG	0x040
+#define DW_DMAC_CHAN_SGR	0x048
+#define DW_DMAC_CHAN_DSR	0x050
+
+#endif /* __AVR32_DW_DMAC_H__ */
Index: linux-2.6.18-avr32/arch/avr32/Kconfig
===================================================================
--- linux-2.6.18-avr32.orig/arch/avr32/Kconfig	2006-11-02 14:17:29.000000000 +0100
+++ linux-2.6.18-avr32/arch/avr32/Kconfig	2006-11-02 15:53:13.000000000 +0100
@@ -157,6 +157,10 @@ config OWNERSHIP_TRACE
 	  enabling Nexus-compliant debuggers to keep track of the PID of the
 	  currently executing task.
 
+config DW_DMAC
+	tristate "Synopsys DesignWare DMA Controller support"
+	default y if CPU_AT32AP7000
+
 # FPU emulation goes here
 
 source "kernel/Kconfig.hz"