diff mbox

[v3,1/2] dma: at_xdmac: creation of the atmel eXtended DMA Controller driver

Message ID 1404825096-15724-2-git-send-email-ludovic.desroches@atmel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Ludovic Desroches July 8, 2014, 1:11 p.m. UTC
New atmel DMA controller known as XDMAC, introduced with SAMA5D4
devices.

Signed-off-by: Ludovic Desroches <ludovic.desroches@atmel.com>
---
 drivers/dma/Kconfig            |    7 +
 drivers/dma/Makefile           |    1 +
 drivers/dma/at_xdmac.c         | 1181 ++++++++++++++++++++++++++++++++++++++++
 drivers/dma/at_xdmac.h         |  308 +++++++++++
 include/dt-bindings/dma/at91.h |   25 +
 5 files changed, 1522 insertions(+)
 create mode 100644 drivers/dma/at_xdmac.c
 create mode 100644 drivers/dma/at_xdmac.h

Comments

Vinod Koul July 25, 2014, 7:16 a.m. UTC | #1
On Tue, Jul 08, 2014 at 03:11:35PM +0200, Ludovic Desroches wrote:
> New atmel DMA controller known as XDMAC, introduced with SAMA5D4
> devices.
Didnt check, but how much is delta b/w X/H Dmacs?

> +/* Call with lock hold. */
> +static void at_xdmac_start_xfer(struct at_xdmac_chan *atchan,
> +				struct at_xdmac_desc *first)
> +{
> +	struct at_xdmac	*atxdmac = to_at_xdmac(atchan->chan.device);
> +	u32		reg;
> +
> +	dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, first);
> +
> +	if (at_xdmac_chan_is_enabled(atchan)) {
> +		dev_err(chan2dev(&atchan->chan),
> +			"BUG: Attempted to start a non-idle channel\n");
This shouldnt be BUG though. Your tasklet is always supposed to start and if
nothing to silent just return!


> +
> +static dma_cookie_t at_xdmac_tx_submit(struct dma_async_tx_descriptor *tx)
> +{
> +	struct at_xdmac_desc	*desc = txd_to_at_desc(tx);
> +	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(tx->chan);
> +	dma_cookie_t		cookie;
> +	unsigned long		flags;
> +
> +	spin_lock_irqsave(&atchan->lock, flags);
> +	cookie = dma_cookie_assign(tx);
> +
> +	dev_vdbg(chan2dev(tx->chan), "%s: atchan 0x%p, add desc 0x%p to xfers_list\n",
> +		 __func__, atchan, desc);
> +	list_add_tail(&desc->xfer_node, &atchan->xfers_list);
> +	if (list_is_singular(&atchan->xfers_list))
> +		at_xdmac_start_xfer(atchan, desc);
so you dont start if list has more than 1 descriptors, why?

> +
> +	spin_unlock_irqrestore(&atchan->lock, flags);
> +	return cookie;
> +}
> +
> +static struct at_xdmac_desc *at_xdmac_alloc_desc(struct dma_chan *chan,
> +						 gfp_t gfp_flags)
why do we need last argument?

> +{
> +	struct at_xdmac_desc	*desc;
> +	struct at_xdmac		*atxdmac = to_at_xdmac(chan->device);
> +	dma_addr_t		phys;
> +
> +	desc = dma_pool_alloc(atxdmac->at_xdmac_desc_pool, gfp_flags, &phys);
> +	if (desc) {
> +		memset(desc, 0, sizeof(*desc));
> +		INIT_LIST_HEAD(&desc->descs_list);
> +		dma_async_tx_descriptor_init(&desc->tx_dma_desc, chan);
> +		desc->tx_dma_desc.tx_submit = at_xdmac_tx_submit;
> +		desc->tx_dma_desc.phys = phys;
> +	}
> +
> +	return desc;
> +}
> +
> +/* Call must be protected by lock. */
> +static struct at_xdmac_desc *at_xdmac_get_desc(struct at_xdmac_chan *atchan)
> +{
> +	struct at_xdmac_desc *desc;
> +
> +	if (list_empty(&atchan->free_descs_list)) {
> +		desc = at_xdmac_alloc_desc(&atchan->chan, GFP_ATOMIC);
GFP_NOWAIT pls

> +static struct dma_async_tx_descriptor *
> +at_xdmac_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
> +			 size_t len, unsigned long flags)
> +{
> +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> +	struct at_xdmac_desc	*first = NULL, *prev = NULL;
> +	size_t			remaining_size = len, xfer_size = 0, ublen;
> +	dma_addr_t		src_addr = src, dst_addr = dest;
> +	u32			dwidth;
> +	/*
> +	 * WARNING: The channel configuration is set here since there is no
> +	 * dmaengine_slave_config call in this case. Moreover we don't know the
> +	 * direction, it involves we can't dynamically set the source and dest
> +	 * interface so we have to use the same one. Only interface 0 allows EBI
> +	 * access. Hopefully we can access DDR through both ports (at least on
> +	 * SAMA5D4x), so we can use the same interface for source and dest,
> +	 * that solves the fact we don't know the direction.
and why do you need slave config for memcpy?

> +static enum dma_status
> +at_xdmac_tx_status(struct dma_chan *chan, dma_cookie_t cookie,
> +		struct dma_tx_state *txstate)
> +{
> +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> +	struct at_xdmac		*atxdmac = to_at_xdmac(atchan->chan.device);
> +	struct at_xdmac_desc	*desc, *_desc;
> +	unsigned long		flags;
> +	enum dma_status		ret;
> +	int			residue;
> +	u32			cur_nda;
> +
> +	ret = dma_cookie_status(chan, cookie, txstate);
> +	if (ret == DMA_COMPLETE)
> +		return ret;
> +
> +	spin_lock_irqsave(&atchan->lock, flags);
> +
> +	desc = list_first_entry(&atchan->xfers_list, struct at_xdmac_desc, xfer_node);
> +
> +	if (!desc->active_xfer)
> +		dev_err(chan2dev(chan),
> +			"something goes wrong, there is no active transfer\n");
We can query resiude even when tx_submit hasnt been invoked. You need to
return the full length in that case.
> +
> +	residue = desc->xfer_size;

Also do check if txsate is NULL, in case just bail out here.

> +
> +	/* Flush FIFO. */
> +	at_xdmac_write(atxdmac, AT_XDMAC_GSWF, atchan->mask);
> +	while (!(at_xdmac_chan_read(atchan, AT_XDMAC_CIS) & AT_XDMAC_CIS_FIS));
> +		cpu_relax();
> +
> +	cur_nda = at_xdmac_chan_read(atchan, AT_XDMAC_CNDA) & 0xfffffffc;
> +	/*
> +	 * Remove size of all microblocks already transferred and the current
> +	 * one. Then add the remaining size to transfer of the current
> +	 * microblock.
> +	 */
> +	list_for_each_entry_safe(desc, _desc, &desc->descs_list, desc_node) {
> +		residue -= (desc->lld.mbr_ubc & 0xffffff) << atchan->dwidth;
> +		if ((desc->lld.mbr_nda & 0xfffffffc) == cur_nda)
> +			break;
> +	}
> +	residue += at_xdmac_chan_read(atchan, AT_XDMAC_CUBC) << atchan->dwidth;
> +
> +	spin_unlock_irqrestore(&atchan->lock, flags);
> +
> +	dma_set_residue(txstate, residue);
> +
> +	dev_dbg(chan2dev(chan),
> +		 "%s: desc=0x%p, tx_dma_desc.phys=0x%08x, tx_status=%d, cookie=%d, residue=%d\n",
> +		 __func__, desc, desc->tx_dma_desc.phys, ret, cookie, residue);
> +
> +	return ret;
> +}
> +
> +static void at_xdmac_terminate_xfer(struct at_xdmac_chan *atchan,
> +				    struct at_xdmac_desc *desc)
> +{
> +	dev_dbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
> +
> +	/*
> +	 * It's necessary to remove the transfer before calling the callback
> +	 * because some devices can call dma_engine_terminate_all causing to do
> +	 * dma_cookie_complete two times on the same cookie.
why would terminate call dma_cookie_complete?

> +	 */
> +	list_del(&desc->xfer_node);
> +	list_splice_init(&desc->descs_list, &atchan->free_descs_list);
shouldnt you move all the desc in active list in one shot ?

> +static void at_xdmac_tasklet(unsigned long data)
> +{
> +	struct at_xdmac_chan	*atchan = (struct at_xdmac_chan *)data;
> +	struct at_xdmac_desc	*desc;
> +	u32			error_mask;
> +
> +	dev_dbg(chan2dev(&atchan->chan), "%s: status=0x%08lx\n",
> +		 __func__, atchan->status);
> +
> +	error_mask = AT_XDMAC_CIS_RBEIS
> +		     | AT_XDMAC_CIS_WBEIS
> +		     | AT_XDMAC_CIS_ROIS;
> +
> +	if (at_xdmac_chan_is_cyclic(atchan)) {
> +		at_xdmac_handle_cyclic(atchan);
> +	} else if ((atchan->status & AT_XDMAC_CIS_LIS)
> +		   || (atchan->status & error_mask)) {
> +		struct dma_async_tx_descriptor  *txd;
> +
> +		if (atchan->status & AT_XDMAC_CIS_RBEIS)
> +			dev_err(chan2dev(&atchan->chan), "read bus error!!!");
> +		else if (atchan->status & AT_XDMAC_CIS_WBEIS)
> +			dev_err(chan2dev(&atchan->chan), "write bus error!!!");
> +		else if (atchan->status & AT_XDMAC_CIS_ROIS)
> +			dev_err(chan2dev(&atchan->chan), "request overflow error!!!");
> +
> +		desc = list_first_entry(&atchan->xfers_list,
> +					struct at_xdmac_desc,
> +					xfer_node);
> +		dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
> +		BUG_ON(!desc->active_xfer);
> +
> +		txd = &desc->tx_dma_desc;
> +
> +		at_xdmac_terminate_xfer(atchan, desc);
why terminate here? This looks wrong

> +static int at_xdmac_control(struct dma_chan *chan, enum dma_ctrl_cmd cmd,
> +			    unsigned long arg)
> +{
> +	struct at_xdmac_desc	*desc, *_desc;
> +	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(chan);
> +	struct at_xdmac		*atxdmac = to_at_xdmac(atchan->chan.device);
> +	unsigned long		flags;
> +	int			ret = 0;
> +
> +	dev_dbg(chan2dev(chan), "%s: cmd=%d\n", __func__, cmd);
> +
> +	spin_lock_irqsave(&atchan->lock, flags);
> +
> +	switch (cmd) {
> +	case DMA_PAUSE:
> +		at_xdmac_write(atxdmac, AT_XDMAC_GRWS, atchan->mask);
> +		set_bit(AT_XDMAC_CHAN_IS_PAUSED, &atchan->status);
> +		break;
empty line here pls

> +	case DMA_RESUME:
> +		if (!at_xdmac_chan_is_paused(atchan))
> +			break;
> +
> +		at_xdmac_write(atxdmac, AT_XDMAC_GRWR, atchan->mask);
> +		clear_bit(AT_XDMAC_CHAN_IS_PAUSED, &atchan->status);
> +		break;
> +	case DMA_TERMINATE_ALL:
> +		at_xdmac_write(atxdmac, AT_XDMAC_GIE, atchan->mask);
> +		at_xdmac_write(atxdmac, AT_XDMAC_GD, atchan->mask);
> +		while (at_xdmac_read(atxdmac, AT_XDMAC_GS) & atchan->mask)
> +			cpu_relax();
> +
> +		/* Cancel all pending transfers. */
> +		list_for_each_entry_safe(desc, _desc, &atchan->xfers_list, xfer_node)
> +			at_xdmac_terminate_xfer(atchan, desc);
> +
> +		clear_bit(AT_XDMAC_CHAN_IS_CYCLIC, &atchan->status);
> +		break;
> +	case DMA_SLAVE_CONFIG:
> +		ret = at_xdmac_set_slave_config(chan,
> +				(struct dma_slave_config *)arg);
> +		break;
> +	default:
> +		dev_err(chan2dev(chan),
> +			"unmanaged or unknown dma control cmd: %d\n", cmd);
> +		ret = -ENXIO;
> +	}
> +
> +	spin_unlock_irqrestore(&atchan->lock, flags);
> +
> +	return ret;
> +}

> +static int atmel_xdmac_resume_noirq(struct device *dev)
> +{
> +	struct platform_device	*pdev = to_platform_device(dev);
> +	struct at_xdmac		*atxdmac = platform_get_drvdata(pdev);
> +	struct at_xdmac_chan	*atchan;
> +	struct dma_chan		*chan, *_chan;
> +	int			i;
> +
> +	clk_prepare_enable(atxdmac->clk);
> +
> +	/* Clear pending interrupts. */
> +	for (i = 0; i < atxdmac->dma.chancnt; i++) {
> +		atchan = &atxdmac->chan[i];
> +		while (at_xdmac_chan_read(atchan, AT_XDMAC_CIS))
> +			cpu_relax();
> +	}
> +
> +	at_xdmac_write(atxdmac, AT_XDMAC_GIE, atxdmac->save_gim);
> +	at_xdmac_write(atxdmac, AT_XDMAC_GE, atxdmac->save_gs);
> +	list_for_each_entry_safe(chan, _chan, &atxdmac->dma.channels, device_node) {
> +		atchan = to_at_xdmac_chan(chan);
> +		at_xdmac_chan_write(atchan, AT_XDMAC_CC, atchan->cfg);
> +		if (at_xdmac_chan_is_cyclic(atchan)) {
> +			at_xdmac_chan_write(atchan, AT_XDMAC_CNDA, atchan->save_cnda);
> +			at_xdmac_chan_write(atchan, AT_XDMAC_CNDC, atchan->save_cndc);
> +			at_xdmac_chan_write(atchan, AT_XDMAC_CIE, atchan->save_cim);
> +			at_xdmac_write(atxdmac, AT_XDMAC_GE, atchan->mask);
> +		}
> +	}
> +	return 0;
> +}
why noirq variants?

> +
> +static int at_xdmac_probe(struct platform_device *pdev)
> +{
> +	struct resource	*res;
> +	struct at_xdmac	*atxdmac;
> +	int		irq, size, nr_channels, i, ret;
> +	void __iomem	*base;
> +	u32		reg;
> +
> +	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
> +	if (!res)
> +		return -EINVAL;
> +
> +	irq = platform_get_irq(pdev, 0);
> +	if (irq < 0)
> +		return irq;
> +
> +	base = devm_ioremap_resource(&pdev->dev, res);
> +	if (IS_ERR(base))
> +		return PTR_ERR(base);
> +
> +	/*
> +	 * Read number of xdmac channels, read helper function can't be used
> +	 * since atxdmac is not yet allocated and we need to know the number
> +	 * of channels to do the allocation.
> +	 */
> +	reg = readl_relaxed(base + AT_XDMAC_GTYPE);
> +	nr_channels = AT_XDMAC_NB_CH(reg);
> +	if (nr_channels > AT_XDMAC_MAX_CHAN) {
> +		dev_err(&pdev->dev, "invalid number of channels (%u)\n",
> +			nr_channels);
> +		return -EINVAL;
> +	}
> +
> +	size = sizeof(*atxdmac);
> +	size += nr_channels * sizeof(struct at_xdmac_chan);
> +	atxdmac = devm_kzalloc(&pdev->dev, size, GFP_KERNEL);
devm_kmalloc_array/devm_kcalloc pls

> +	if (!atxdmac) {
> +		dev_err(&pdev->dev, "can't allocate at_xdmac structure\n");
> +		return -ENOMEM;
> +	}
> +
> +	atxdmac->regs = base;
> +	atxdmac->irq = irq;
> +
> +	ret = devm_request_irq(&pdev->dev, atxdmac->irq, at_xdmac_interrupt, 0,
> +			       "at_xdmac", atxdmac);
and this can invoke your irq handler and you may not be ready. Second this
cause races with tasklet, so usualy recommednation is to use regular
request_irq()
> +	if (ret) {
> +		dev_err(&pdev->dev, "can't request irq\n");
> +		return ret;
> +	}
> +
> +	atxdmac->clk = devm_clk_get(&pdev->dev, "dma_clk");
> +	if (IS_ERR(atxdmac->clk)) {
> +		dev_err(&pdev->dev, "can't get dma_clk\n");
> +		return PTR_ERR(atxdmac->clk);
> +	}
> +
> +	ret = clk_prepare_enable(atxdmac->clk);
> +	if (ret) {
> +		dev_err(&pdev->dev, "can't prepare or enable clock\n");
> +		return ret;
> +	}
> +
> +	atxdmac->at_xdmac_desc_pool =
> +		dmam_pool_create(dev_name(&pdev->dev), &pdev->dev,
> +				sizeof(struct at_xdmac_desc), 4, 0);
> +	if (!atxdmac->at_xdmac_desc_pool) {
> +		dev_err(&pdev->dev, "no memory for descriptors dma pool\n");
> +		ret = -ENOMEM;
> +		goto err_clk_disable;
> +	}
> +
> +	dma_cap_set(DMA_CYCLIC, atxdmac->dma.cap_mask);
> +	dma_cap_set(DMA_MEMCPY, atxdmac->dma.cap_mask);
> +	dma_cap_set(DMA_SLAVE, atxdmac->dma.cap_mask);
> +	atxdmac->dma.dev				= &pdev->dev;
> +	atxdmac->dma.device_alloc_chan_resources	= at_xdmac_alloc_chan_resources;
> +	atxdmac->dma.device_free_chan_resources		= at_xdmac_free_chan_resources;
> +	atxdmac->dma.device_tx_status			= at_xdmac_tx_status;
> +	atxdmac->dma.device_issue_pending		= at_xdmac_issue_pending;
> +	atxdmac->dma.device_prep_dma_cyclic		= at_xdmac_prep_dma_cyclic;
> +	atxdmac->dma.device_prep_dma_memcpy		= at_xdmac_prep_dma_memcpy;
> +	atxdmac->dma.device_prep_slave_sg		= at_xdmac_prep_slave_sg;
> +	atxdmac->dma.device_control			= at_xdmac_control;
> +	atxdmac->dma.chancnt				= nr_channels;
no caps, pls do implement that as you have cyclic dma too
> +
> +	/* Disable all chans and interrupts. */
> +	at_xdmac_off(atxdmac);
> +
> +	/* Init channels. */
> +	INIT_LIST_HEAD(&atxdmac->dma.channels);
> +	for (i = 0; i < nr_channels; i++) {
> +		struct at_xdmac_chan *atchan = &atxdmac->chan[i];
> +
> +		atchan->chan.device = &atxdmac->dma;
> +		list_add_tail(&atchan->chan.device_node,
> +			      &atxdmac->dma.channels);
> +
> +		atchan->ch_regs = at_xdmac_chan_reg_base(atxdmac, i);
> +		atchan->mask = 1 << i;
> +
> +		spin_lock_init(&atchan->lock);
> +		INIT_LIST_HEAD(&atchan->xfers_list);
> +		INIT_LIST_HEAD(&atchan->free_descs_list);
> +		tasklet_init(&atchan->tasklet, at_xdmac_tasklet,
> +			     (unsigned long)atchan);
> +
> +		/* Clear pending interrupts. */
> +		while (at_xdmac_chan_read(atchan, AT_XDMAC_CIS))
> +			cpu_relax();
> +	}
> +	platform_set_drvdata(pdev, atxdmac);
> +
> +	ret = dma_async_device_register(&atxdmac->dma);
> +	if (ret) {
> +		dev_err(&pdev->dev, "fail to register DMA engine device\n");
> +		goto err_clk_disable;
> +	}
> +
> +	ret = of_dma_controller_register(pdev->dev.of_node,
> +					 at_xdmac_xlate, atxdmac);
> +	if (ret) {
> +		dev_err(&pdev->dev, "could not register of dma controller\n");
> +		goto err_dma_unregister;
> +	}
> +
> +	dev_info(&pdev->dev, "%d channels, mapped at 0x%p\n",
> +		 nr_channels, atxdmac->regs);
> +
> +	return 0;
> +
> +err_dma_unregister:
> +	dma_async_device_unregister(&atxdmac->dma);
> +err_clk_disable:
> +	clk_disable_unprepare(atxdmac->clk);
> +	return ret;
> +}
> +
> +static int at_xdmac_remove(struct platform_device *pdev)
> +{
> +	struct at_xdmac	*atxdmac = (struct at_xdmac*)platform_get_drvdata(pdev);
> +	int		i;
> +
> +	at_xdmac_off(atxdmac);
> +	of_dma_controller_free(pdev->dev.of_node);
> +	dma_async_device_unregister(&atxdmac->dma);
> +	clk_disable_unprepare(atxdmac->clk);
> +
> +	synchronize_irq(atxdmac->irq);
> +
> +	for (i = 0; i < atxdmac->dma.chancnt; i++) {
> +		struct at_xdmac_chan *atchan = &atxdmac->chan[i];
> +
> +		tasklet_kill(&atchan->tasklet);
> +		at_xdmac_free_chan_resources(&atchan->chan);
> +	}
and you can still get irq!
> +
> +	return 0;
> +}
> +
> +static const struct dev_pm_ops atmel_xdmac_dev_pm_ops = {
> +	.prepare	= atmel_xdmac_prepare,
> +	.suspend_noirq	= atmel_xdmac_suspend_noirq,
> +	.resume_noirq	= atmel_xdmac_resume_noirq,
> +};
no ifdef CONFIG_PM?
You might want to make them late_suspend and early_resume to allow
clients suspending first

Thanks
Ludovic Desroches July 28, 2014, 3:54 p.m. UTC | #2
Hello Vinod,

Thanks for your review,

On Fri, Jul 25, 2014 at 12:46:31PM +0530, Vinod Koul wrote:
> On Tue, Jul 08, 2014 at 03:11:35PM +0200, Ludovic Desroches wrote:
> > New atmel DMA controller known as XDMAC, introduced with SAMA5D4
> > devices.
> Didnt check, but how much is delta b/w X/H Dmacs?

If you are wondering if we can have only one driver for both DMA
controller, I don't think so. At the moment you won't find documentation
about this controller since the device is not yet released.

The content of the register has changed, the way irq is managed has
changed too.

> 
> > +/* Call with lock hold. */
> > +static void at_xdmac_start_xfer(struct at_xdmac_chan *atchan,
> > +				struct at_xdmac_desc *first)
> > +{
> > +	struct at_xdmac	*atxdmac = to_at_xdmac(atchan->chan.device);
> > +	u32		reg;
> > +
> > +	dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, first);
> > +
> > +	if (at_xdmac_chan_is_enabled(atchan)) {
> > +		dev_err(chan2dev(&atchan->chan),
> > +			"BUG: Attempted to start a non-idle channel\n");
> This shouldnt be BUG though. Your tasklet is always supposed to start and if
> nothing to silent just return!
> 

Ok, I'll remove this error message.

> 
> > +
> > +static dma_cookie_t at_xdmac_tx_submit(struct dma_async_tx_descriptor *tx)
> > +{
> > +	struct at_xdmac_desc	*desc = txd_to_at_desc(tx);
> > +	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(tx->chan);
> > +	dma_cookie_t		cookie;
> > +	unsigned long		flags;
> > +
> > +	spin_lock_irqsave(&atchan->lock, flags);
> > +	cookie = dma_cookie_assign(tx);
> > +
> > +	dev_vdbg(chan2dev(tx->chan), "%s: atchan 0x%p, add desc 0x%p to xfers_list\n",
> > +		 __func__, atchan, desc);
> > +	list_add_tail(&desc->xfer_node, &atchan->xfers_list);
> > +	if (list_is_singular(&atchan->xfers_list))
> > +		at_xdmac_start_xfer(atchan, desc);
> so you dont start if list has more than 1 descriptors, why?
> 

Because I will let at_xdmac_advance_work managing the queue if not
empty. So the only moment when I have to start the transfer in tx_submit
is when I have only one transfer in the queue.

> > +
> > +	spin_unlock_irqrestore(&atchan->lock, flags);
> > +	return cookie;
> > +}
> > +
> > +static struct at_xdmac_desc *at_xdmac_alloc_desc(struct dma_chan *chan,
> > +						 gfp_t gfp_flags)
> why do we need last argument?
> 

No specific reason, I can remove it and use GFP_NOWAIT directly.

> > +{
> > +	struct at_xdmac_desc	*desc;
> > +	struct at_xdmac		*atxdmac = to_at_xdmac(chan->device);
> > +	dma_addr_t		phys;
> > +
> > +	desc = dma_pool_alloc(atxdmac->at_xdmac_desc_pool, gfp_flags, &phys);
> > +	if (desc) {
> > +		memset(desc, 0, sizeof(*desc));
> > +		INIT_LIST_HEAD(&desc->descs_list);
> > +		dma_async_tx_descriptor_init(&desc->tx_dma_desc, chan);
> > +		desc->tx_dma_desc.tx_submit = at_xdmac_tx_submit;
> > +		desc->tx_dma_desc.phys = phys;
> > +	}
> > +
> > +	return desc;
> > +}
> > +
> > +/* Call must be protected by lock. */
> > +static struct at_xdmac_desc *at_xdmac_get_desc(struct at_xdmac_chan *atchan)
> > +{
> > +	struct at_xdmac_desc *desc;
> > +
> > +	if (list_empty(&atchan->free_descs_list)) {
> > +		desc = at_xdmac_alloc_desc(&atchan->chan, GFP_ATOMIC);
> GFP_NOWAIT pls
> 

Ok

> > +static struct dma_async_tx_descriptor *
> > +at_xdmac_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
> > +			 size_t len, unsigned long flags)
> > +{
> > +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> > +	struct at_xdmac_desc	*first = NULL, *prev = NULL;
> > +	size_t			remaining_size = len, xfer_size = 0, ublen;
> > +	dma_addr_t		src_addr = src, dst_addr = dest;
> > +	u32			dwidth;
> > +	/*
> > +	 * WARNING: The channel configuration is set here since there is no
> > +	 * dmaengine_slave_config call in this case. Moreover we don't know the
> > +	 * direction, it involves we can't dynamically set the source and dest
> > +	 * interface so we have to use the same one. Only interface 0 allows EBI
> > +	 * access. Hopefully we can access DDR through both ports (at least on
> > +	 * SAMA5D4x), so we can use the same interface for source and dest,
> > +	 * that solves the fact we don't know the direction.
> and why do you need slave config for memcpy?
> 

Maybe the comments is not clear. With slave config we have the direction
per to mem or mem to per. Of course when doing memcpy we don't need this
information. But I use this information not only to set the transfer
direction but also to set the interfaces to use for the source and dest
(these interfaces depend on the connection on the matrix).
So slave config was useful due to direction field  to tell which interface
to use as source and which one as dest.
I am lucky because NAND and DDR are on the same interfaces so it's not a
problem.

This comment is more a warning for myself to keep in mind this possible 
issue in order to not have a device with NAND and DDR on different
interfaces because it will be difficult to tell which interface is the
source and which one is the dest.

> > +static enum dma_status
> > +at_xdmac_tx_status(struct dma_chan *chan, dma_cookie_t cookie,
> > +		struct dma_tx_state *txstate)
> > +{
> > +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> > +	struct at_xdmac		*atxdmac = to_at_xdmac(atchan->chan.device);
> > +	struct at_xdmac_desc	*desc, *_desc;
> > +	unsigned long		flags;
> > +	enum dma_status		ret;
> > +	int			residue;
> > +	u32			cur_nda;
> > +
> > +	ret = dma_cookie_status(chan, cookie, txstate);
> > +	if (ret == DMA_COMPLETE)
> > +		return ret;
> > +
> > +	spin_lock_irqsave(&atchan->lock, flags);
> > +
> > +	desc = list_first_entry(&atchan->xfers_list, struct at_xdmac_desc, xfer_node);
> > +
> > +	if (!desc->active_xfer)
> > +		dev_err(chan2dev(chan),
> > +			"something goes wrong, there is no active transfer\n");
> We can query resiude even when tx_submit hasnt been invoked. You need to
> return the full length in that case.

Ok, I was not aware about this case.

> > +
> > +	residue = desc->xfer_size;
> 
> Also do check if txsate is NULL, in case just bail out here.

Return value will be the one returned by dma_cookie_status?

> 
> > +
> > +	/* Flush FIFO. */
> > +	at_xdmac_write(atxdmac, AT_XDMAC_GSWF, atchan->mask);
> > +	while (!(at_xdmac_chan_read(atchan, AT_XDMAC_CIS) & AT_XDMAC_CIS_FIS));
> > +		cpu_relax();
> > +
> > +	cur_nda = at_xdmac_chan_read(atchan, AT_XDMAC_CNDA) & 0xfffffffc;
> > +	/*
> > +	 * Remove size of all microblocks already transferred and the current
> > +	 * one. Then add the remaining size to transfer of the current
> > +	 * microblock.
> > +	 */
> > +	list_for_each_entry_safe(desc, _desc, &desc->descs_list, desc_node) {
> > +		residue -= (desc->lld.mbr_ubc & 0xffffff) << atchan->dwidth;
> > +		if ((desc->lld.mbr_nda & 0xfffffffc) == cur_nda)
> > +			break;
> > +	}
> > +	residue += at_xdmac_chan_read(atchan, AT_XDMAC_CUBC) << atchan->dwidth;
> > +
> > +	spin_unlock_irqrestore(&atchan->lock, flags);
> > +
> > +	dma_set_residue(txstate, residue);
> > +
> > +	dev_dbg(chan2dev(chan),
> > +		 "%s: desc=0x%p, tx_dma_desc.phys=0x%08x, tx_status=%d, cookie=%d, residue=%d\n",
> > +		 __func__, desc, desc->tx_dma_desc.phys, ret, cookie, residue);
> > +
> > +	return ret;
> > +}
> > +
> > +static void at_xdmac_terminate_xfer(struct at_xdmac_chan *atchan,
> > +				    struct at_xdmac_desc *desc)
> > +{
> > +	dev_dbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
> > +
> > +	/*
> > +	 * It's necessary to remove the transfer before calling the callback
> > +	 * because some devices can call dma_engine_terminate_all causing to do
> > +	 * dma_cookie_complete two times on the same cookie.
> why would terminate call dma_cookie_complete?
> 

This comment should be remove. It was a comment from the first draft I
did, I have forgotten to remove it. At that time I was calling
dma_cookie_complete here.

> > +	 */
> > +	list_del(&desc->xfer_node);
> > +	list_splice_init(&desc->descs_list, &atchan->free_descs_list);
> shouldnt you move all the desc in active list in one shot ?
> 

Sorry I don't get it. What are you suggesting by one shot?

> > +static void at_xdmac_tasklet(unsigned long data)
> > +{
> > +	struct at_xdmac_chan	*atchan = (struct at_xdmac_chan *)data;
> > +	struct at_xdmac_desc	*desc;
> > +	u32			error_mask;
> > +
> > +	dev_dbg(chan2dev(&atchan->chan), "%s: status=0x%08lx\n",
> > +		 __func__, atchan->status);
> > +
> > +	error_mask = AT_XDMAC_CIS_RBEIS
> > +		     | AT_XDMAC_CIS_WBEIS
> > +		     | AT_XDMAC_CIS_ROIS;
> > +
> > +	if (at_xdmac_chan_is_cyclic(atchan)) {
> > +		at_xdmac_handle_cyclic(atchan);
> > +	} else if ((atchan->status & AT_XDMAC_CIS_LIS)
> > +		   || (atchan->status & error_mask)) {
> > +		struct dma_async_tx_descriptor  *txd;
> > +
> > +		if (atchan->status & AT_XDMAC_CIS_RBEIS)
> > +			dev_err(chan2dev(&atchan->chan), "read bus error!!!");
> > +		else if (atchan->status & AT_XDMAC_CIS_WBEIS)
> > +			dev_err(chan2dev(&atchan->chan), "write bus error!!!");
> > +		else if (atchan->status & AT_XDMAC_CIS_ROIS)
> > +			dev_err(chan2dev(&atchan->chan), "request overflow error!!!");
> > +
> > +		desc = list_first_entry(&atchan->xfers_list,
> > +					struct at_xdmac_desc,
> > +					xfer_node);
> > +		dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
> > +		BUG_ON(!desc->active_xfer);
> > +
> > +		txd = &desc->tx_dma_desc;
> > +
> > +		at_xdmac_terminate_xfer(atchan, desc);
> why terminate here? This looks wrong

Why? What do you have in mind? It is not obvious for me that it is the
wrong place.

> 
> > +static int at_xdmac_control(struct dma_chan *chan, enum dma_ctrl_cmd cmd,
> > +			    unsigned long arg)
> > +{
> > +	struct at_xdmac_desc	*desc, *_desc;
> > +	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(chan);
> > +	struct at_xdmac		*atxdmac = to_at_xdmac(atchan->chan.device);
> > +	unsigned long		flags;
> > +	int			ret = 0;
> > +
> > +	dev_dbg(chan2dev(chan), "%s: cmd=%d\n", __func__, cmd);
> > +
> > +	spin_lock_irqsave(&atchan->lock, flags);
> > +
> > +	switch (cmd) {
> > +	case DMA_PAUSE:
> > +		at_xdmac_write(atxdmac, AT_XDMAC_GRWS, atchan->mask);
> > +		set_bit(AT_XDMAC_CHAN_IS_PAUSED, &atchan->status);
> > +		break;
> empty line here pls
>

Ok

> > +	case DMA_RESUME:
> > +		if (!at_xdmac_chan_is_paused(atchan))
> > +			break;
> > +
> > +		at_xdmac_write(atxdmac, AT_XDMAC_GRWR, atchan->mask);
> > +		clear_bit(AT_XDMAC_CHAN_IS_PAUSED, &atchan->status);
> > +		break;
> > +	case DMA_TERMINATE_ALL:
> > +		at_xdmac_write(atxdmac, AT_XDMAC_GIE, atchan->mask);
> > +		at_xdmac_write(atxdmac, AT_XDMAC_GD, atchan->mask);
> > +		while (at_xdmac_read(atxdmac, AT_XDMAC_GS) & atchan->mask)
> > +			cpu_relax();
> > +
> > +		/* Cancel all pending transfers. */
> > +		list_for_each_entry_safe(desc, _desc, &atchan->xfers_list, xfer_node)
> > +			at_xdmac_terminate_xfer(atchan, desc);
> > +
> > +		clear_bit(AT_XDMAC_CHAN_IS_CYCLIC, &atchan->status);
> > +		break;
> > +	case DMA_SLAVE_CONFIG:
> > +		ret = at_xdmac_set_slave_config(chan,
> > +				(struct dma_slave_config *)arg);
> > +		break;
> > +	default:
> > +		dev_err(chan2dev(chan),
> > +			"unmanaged or unknown dma control cmd: %d\n", cmd);
> > +		ret = -ENXIO;
> > +	}
> > +
> > +	spin_unlock_irqrestore(&atchan->lock, flags);
> > +
> > +	return ret;
> > +}
> 
> > +static int atmel_xdmac_resume_noirq(struct device *dev)
> > +{
> > +	struct platform_device	*pdev = to_platform_device(dev);
> > +	struct at_xdmac		*atxdmac = platform_get_drvdata(pdev);
> > +	struct at_xdmac_chan	*atchan;
> > +	struct dma_chan		*chan, *_chan;
> > +	int			i;
> > +
> > +	clk_prepare_enable(atxdmac->clk);
> > +
> > +	/* Clear pending interrupts. */
> > +	for (i = 0; i < atxdmac->dma.chancnt; i++) {
> > +		atchan = &atxdmac->chan[i];
> > +		while (at_xdmac_chan_read(atchan, AT_XDMAC_CIS))
> > +			cpu_relax();
> > +	}
> > +
> > +	at_xdmac_write(atxdmac, AT_XDMAC_GIE, atxdmac->save_gim);
> > +	at_xdmac_write(atxdmac, AT_XDMAC_GE, atxdmac->save_gs);
> > +	list_for_each_entry_safe(chan, _chan, &atxdmac->dma.channels, device_node) {
> > +		atchan = to_at_xdmac_chan(chan);
> > +		at_xdmac_chan_write(atchan, AT_XDMAC_CC, atchan->cfg);
> > +		if (at_xdmac_chan_is_cyclic(atchan)) {
> > +			at_xdmac_chan_write(atchan, AT_XDMAC_CNDA, atchan->save_cnda);
> > +			at_xdmac_chan_write(atchan, AT_XDMAC_CNDC, atchan->save_cndc);
> > +			at_xdmac_chan_write(atchan, AT_XDMAC_CIE, atchan->save_cim);
> > +			at_xdmac_write(atxdmac, AT_XDMAC_GE, atchan->mask);
> > +		}
> > +	}
> > +	return 0;
> > +}
> why noirq variants?

I was wondering which variants should be the one I have to use. Since I
was not sure, I have chosen to do it in the same way as at_hdmac driver.

> 
> > +
> > +static int at_xdmac_probe(struct platform_device *pdev)
> > +{
> > +	struct resource	*res;
> > +	struct at_xdmac	*atxdmac;
> > +	int		irq, size, nr_channels, i, ret;
> > +	void __iomem	*base;
> > +	u32		reg;
> > +
> > +	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
> > +	if (!res)
> > +		return -EINVAL;
> > +
> > +	irq = platform_get_irq(pdev, 0);
> > +	if (irq < 0)
> > +		return irq;
> > +
> > +	base = devm_ioremap_resource(&pdev->dev, res);
> > +	if (IS_ERR(base))
> > +		return PTR_ERR(base);
> > +
> > +	/*
> > +	 * Read number of xdmac channels, read helper function can't be used
> > +	 * since atxdmac is not yet allocated and we need to know the number
> > +	 * of channels to do the allocation.
> > +	 */
> > +	reg = readl_relaxed(base + AT_XDMAC_GTYPE);
> > +	nr_channels = AT_XDMAC_NB_CH(reg);
> > +	if (nr_channels > AT_XDMAC_MAX_CHAN) {
> > +		dev_err(&pdev->dev, "invalid number of channels (%u)\n",
> > +			nr_channels);
> > +		return -EINVAL;
> > +	}
> > +
> > +	size = sizeof(*atxdmac);
> > +	size += nr_channels * sizeof(struct at_xdmac_chan);
> > +	atxdmac = devm_kzalloc(&pdev->dev, size, GFP_KERNEL);
> devm_kmalloc_array/devm_kcalloc pls

Useful for allocating the channels structures but there is also the
at_xdmac structure.

> 
> > +	if (!atxdmac) {
> > +		dev_err(&pdev->dev, "can't allocate at_xdmac structure\n");
> > +		return -ENOMEM;
> > +	}
> > +
> > +	atxdmac->regs = base;
> > +	atxdmac->irq = irq;
> > +
> > +	ret = devm_request_irq(&pdev->dev, atxdmac->irq, at_xdmac_interrupt, 0,
> > +			       "at_xdmac", atxdmac);
> and this can invoke your irq handler and you may not be ready. Second this
> cause races with tasklet, so usualy recommednation is to use regular
> request_irq()

You mean it is specific to devm variant, isn't it?

> > +	if (ret) {
> > +		dev_err(&pdev->dev, "can't request irq\n");
> > +		return ret;
> > +	}
> > +
> > +	atxdmac->clk = devm_clk_get(&pdev->dev, "dma_clk");
> > +	if (IS_ERR(atxdmac->clk)) {
> > +		dev_err(&pdev->dev, "can't get dma_clk\n");
> > +		return PTR_ERR(atxdmac->clk);
> > +	}
> > +
> > +	ret = clk_prepare_enable(atxdmac->clk);
> > +	if (ret) {
> > +		dev_err(&pdev->dev, "can't prepare or enable clock\n");
> > +		return ret;
> > +	}
> > +
> > +	atxdmac->at_xdmac_desc_pool =
> > +		dmam_pool_create(dev_name(&pdev->dev), &pdev->dev,
> > +				sizeof(struct at_xdmac_desc), 4, 0);
> > +	if (!atxdmac->at_xdmac_desc_pool) {
> > +		dev_err(&pdev->dev, "no memory for descriptors dma pool\n");
> > +		ret = -ENOMEM;
> > +		goto err_clk_disable;
> > +	}
> > +
> > +	dma_cap_set(DMA_CYCLIC, atxdmac->dma.cap_mask);
> > +	dma_cap_set(DMA_MEMCPY, atxdmac->dma.cap_mask);
> > +	dma_cap_set(DMA_SLAVE, atxdmac->dma.cap_mask);
> > +	atxdmac->dma.dev				= &pdev->dev;
> > +	atxdmac->dma.device_alloc_chan_resources	= at_xdmac_alloc_chan_resources;
> > +	atxdmac->dma.device_free_chan_resources		= at_xdmac_free_chan_resources;
> > +	atxdmac->dma.device_tx_status			= at_xdmac_tx_status;
> > +	atxdmac->dma.device_issue_pending		= at_xdmac_issue_pending;
> > +	atxdmac->dma.device_prep_dma_cyclic		= at_xdmac_prep_dma_cyclic;
> > +	atxdmac->dma.device_prep_dma_memcpy		= at_xdmac_prep_dma_memcpy;
> > +	atxdmac->dma.device_prep_slave_sg		= at_xdmac_prep_slave_sg;
> > +	atxdmac->dma.device_control			= at_xdmac_control;
> > +	atxdmac->dma.chancnt				= nr_channels;
> no caps, pls do implement that as you have cyclic dma too

Sorry I don't understand what you mean here.

> > +
> > +	/* Disable all chans and interrupts. */
> > +	at_xdmac_off(atxdmac);
> > +
> > +	/* Init channels. */
> > +	INIT_LIST_HEAD(&atxdmac->dma.channels);
> > +	for (i = 0; i < nr_channels; i++) {
> > +		struct at_xdmac_chan *atchan = &atxdmac->chan[i];
> > +
> > +		atchan->chan.device = &atxdmac->dma;
> > +		list_add_tail(&atchan->chan.device_node,
> > +			      &atxdmac->dma.channels);
> > +
> > +		atchan->ch_regs = at_xdmac_chan_reg_base(atxdmac, i);
> > +		atchan->mask = 1 << i;
> > +
> > +		spin_lock_init(&atchan->lock);
> > +		INIT_LIST_HEAD(&atchan->xfers_list);
> > +		INIT_LIST_HEAD(&atchan->free_descs_list);
> > +		tasklet_init(&atchan->tasklet, at_xdmac_tasklet,
> > +			     (unsigned long)atchan);
> > +
> > +		/* Clear pending interrupts. */
> > +		while (at_xdmac_chan_read(atchan, AT_XDMAC_CIS))
> > +			cpu_relax();
> > +	}
> > +	platform_set_drvdata(pdev, atxdmac);
> > +
> > +	ret = dma_async_device_register(&atxdmac->dma);
> > +	if (ret) {
> > +		dev_err(&pdev->dev, "fail to register DMA engine device\n");
> > +		goto err_clk_disable;
> > +	}
> > +
> > +	ret = of_dma_controller_register(pdev->dev.of_node,
> > +					 at_xdmac_xlate, atxdmac);
> > +	if (ret) {
> > +		dev_err(&pdev->dev, "could not register of dma controller\n");
> > +		goto err_dma_unregister;
> > +	}
> > +
> > +	dev_info(&pdev->dev, "%d channels, mapped at 0x%p\n",
> > +		 nr_channels, atxdmac->regs);
> > +
> > +	return 0;
> > +
> > +err_dma_unregister:
> > +	dma_async_device_unregister(&atxdmac->dma);
> > +err_clk_disable:
> > +	clk_disable_unprepare(atxdmac->clk);
> > +	return ret;
> > +}
> > +
> > +static int at_xdmac_remove(struct platform_device *pdev)
> > +{
> > +	struct at_xdmac	*atxdmac = (struct at_xdmac*)platform_get_drvdata(pdev);
> > +	int		i;
> > +
> > +	at_xdmac_off(atxdmac);
> > +	of_dma_controller_free(pdev->dev.of_node);
> > +	dma_async_device_unregister(&atxdmac->dma);
> > +	clk_disable_unprepare(atxdmac->clk);
> > +
> > +	synchronize_irq(atxdmac->irq);
> > +
> > +	for (i = 0; i < atxdmac->dma.chancnt; i++) {
> > +		struct at_xdmac_chan *atchan = &atxdmac->chan[i];
> > +
> > +		tasklet_kill(&atchan->tasklet);
> > +		at_xdmac_free_chan_resources(&atchan->chan);
> > +	}
> and you can still get irq!

Why? I thought deactivating irq on device side and calling synchronize
irq will do the stuff. 

> > +
> > +	return 0;
> > +}
> > +
> > +static const struct dev_pm_ops atmel_xdmac_dev_pm_ops = {
> > +	.prepare	= atmel_xdmac_prepare,
> > +	.suspend_noirq	= atmel_xdmac_suspend_noirq,
> > +	.resume_noirq	= atmel_xdmac_resume_noirq,
> > +};
> no ifdef CONFIG_PM?

I'll add it.

> You might want to make them late_suspend and early_resume to allow
> clients suspending first

Ok, as I told before I was wondering which variant to use. For my
knowledge, it will be the case also with noirq, isn't it? It is called
after suspend_late.

> 
> Thanks
> -- 
> ~Vinod
> --
> To unsubscribe from this list: send the line "unsubscribe dmaengine" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
Vinod Koul July 28, 2014, 4:47 p.m. UTC | #3
On Mon, Jul 28, 2014 at 05:54:31PM +0200, Ludovic Desroches wrote:
> > > +{
> > > +	struct at_xdmac_desc	*desc = txd_to_at_desc(tx);
> > > +	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(tx->chan);
> > > +	dma_cookie_t		cookie;
> > > +	unsigned long		flags;
> > > +
> > > +	spin_lock_irqsave(&atchan->lock, flags);
> > > +	cookie = dma_cookie_assign(tx);
> > > +
> > > +	dev_vdbg(chan2dev(tx->chan), "%s: atchan 0x%p, add desc 0x%p to xfers_list\n",
> > > +		 __func__, atchan, desc);
> > > +	list_add_tail(&desc->xfer_node, &atchan->xfers_list);
> > > +	if (list_is_singular(&atchan->xfers_list))
> > > +		at_xdmac_start_xfer(atchan, desc);
> > so you dont start if list has more than 1 descriptors, why?
> Because I will let at_xdmac_advance_work managing the queue if not
> empty. So the only moment when I have to start the transfer in tx_submit
> is when I have only one transfer in the queue.
So xfers_list is current active descriptors right and if it is always going
to be singular why bother with a list?

> > > +static struct dma_async_tx_descriptor *
> > > +at_xdmac_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
> > > +			 size_t len, unsigned long flags)
> > > +{
> > > +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> > > +	struct at_xdmac_desc	*first = NULL, *prev = NULL;
> > > +	size_t			remaining_size = len, xfer_size = 0, ublen;
> > > +	dma_addr_t		src_addr = src, dst_addr = dest;
> > > +	u32			dwidth;
> > > +	/*
> > > +	 * WARNING: The channel configuration is set here since there is no
> > > +	 * dmaengine_slave_config call in this case. Moreover we don't know the
> > > +	 * direction, it involves we can't dynamically set the source and dest
> > > +	 * interface so we have to use the same one. Only interface 0 allows EBI
> > > +	 * access. Hopefully we can access DDR through both ports (at least on
> > > +	 * SAMA5D4x), so we can use the same interface for source and dest,
> > > +	 * that solves the fact we don't know the direction.
> > and why do you need slave config for memcpy?
> Maybe the comments is not clear. With slave config we have the direction
> per to mem or mem to per. Of course when doing memcpy we don't need this
> information. But I use this information not only to set the transfer
> direction but also to set the interfaces to use for the source and dest
> (these interfaces depend on the connection on the matrix).
> So slave config was useful due to direction field  to tell which interface
> to use as source and which one as dest.
> I am lucky because NAND and DDR are on the same interfaces so it's not a
> problem.
> 
> This comment is more a warning for myself to keep in mind this possible 
> issue in order to not have a device with NAND and DDR on different
> interfaces because it will be difficult to tell which interface is the
> source and which one is the dest.
Well in that case shouldnt NAND be treated like periphral and not memory?

> > > +static enum dma_status
> > > +at_xdmac_tx_status(struct dma_chan *chan, dma_cookie_t cookie,
> > > +		struct dma_tx_state *txstate)
> > > +{
> > > +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> > > +	struct at_xdmac		*atxdmac = to_at_xdmac(atchan->chan.device);
> > > +	struct at_xdmac_desc	*desc, *_desc;
> > > +	unsigned long		flags;
> > > +	enum dma_status		ret;
> > > +	int			residue;
> > > +	u32			cur_nda;
> > > +
> > > +	ret = dma_cookie_status(chan, cookie, txstate);
> > > +	if (ret == DMA_COMPLETE)
> > > +		return ret;
> > > +
> > > +	spin_lock_irqsave(&atchan->lock, flags);
> > > +
> > > +	desc = list_first_entry(&atchan->xfers_list, struct at_xdmac_desc, xfer_node);
> > > +
> > > +	if (!desc->active_xfer)
> > > +		dev_err(chan2dev(chan),
> > > +			"something goes wrong, there is no active transfer\n");
> > We can query resiude even when tx_submit hasnt been invoked. You need to
> > return the full length in that case.
> 
> Ok, I was not aware about this case.
> 
> > > +
> > > +	residue = desc->xfer_size;
> > 
> > Also do check if txsate is NULL, in case just bail out here.
> 
> Return value will be the one returned by dma_cookie_status?
Yes

> > > +	 */
> > > +	list_del(&desc->xfer_node);
> > > +	list_splice_init(&desc->descs_list, &atchan->free_descs_list);
> > shouldnt you move all the desc in active list in one shot ?
> > 
> 
> Sorry I don't get it. What are you suggesting by one shot?
Rather than invoking this per descriptor moving descriptors to
free_desc_list one by one, we can move all descriptors togther.

> 
> > > +static void at_xdmac_tasklet(unsigned long data)
> > > +{
> > > +	struct at_xdmac_chan	*atchan = (struct at_xdmac_chan *)data;
> > > +	struct at_xdmac_desc	*desc;
> > > +	u32			error_mask;
> > > +
> > > +	dev_dbg(chan2dev(&atchan->chan), "%s: status=0x%08lx\n",
> > > +		 __func__, atchan->status);
> > > +
> > > +	error_mask = AT_XDMAC_CIS_RBEIS
> > > +		     | AT_XDMAC_CIS_WBEIS
> > > +		     | AT_XDMAC_CIS_ROIS;
> > > +
> > > +	if (at_xdmac_chan_is_cyclic(atchan)) {
> > > +		at_xdmac_handle_cyclic(atchan);
> > > +	} else if ((atchan->status & AT_XDMAC_CIS_LIS)
> > > +		   || (atchan->status & error_mask)) {
> > > +		struct dma_async_tx_descriptor  *txd;
> > > +
> > > +		if (atchan->status & AT_XDMAC_CIS_RBEIS)
> > > +			dev_err(chan2dev(&atchan->chan), "read bus error!!!");
> > > +		else if (atchan->status & AT_XDMAC_CIS_WBEIS)
> > > +			dev_err(chan2dev(&atchan->chan), "write bus error!!!");
> > > +		else if (atchan->status & AT_XDMAC_CIS_ROIS)
> > > +			dev_err(chan2dev(&atchan->chan), "request overflow error!!!");
> > > +
> > > +		desc = list_first_entry(&atchan->xfers_list,
> > > +					struct at_xdmac_desc,
> > > +					xfer_node);
> > > +		dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
> > > +		BUG_ON(!desc->active_xfer);
> > > +
> > > +		txd = &desc->tx_dma_desc;
> > > +
> > > +		at_xdmac_terminate_xfer(atchan, desc);
> > why terminate here? This looks wrong
> 
> Why? What do you have in mind? It is not obvious for me that it is the
> wrong place.
The descriptor has completed. The terminate has special meaning. The user
can cancel all transactions by invoking terminate_all(). That implies to
clean up the channel and abort the transfers. The transfer is done so what
does it mean to terminate.

> > > +	if (!atxdmac) {
> > > +		dev_err(&pdev->dev, "can't allocate at_xdmac structure\n");
> > > +		return -ENOMEM;
> > > +	}
> > > +
> > > +	atxdmac->regs = base;
> > > +	atxdmac->irq = irq;
> > > +
> > > +	ret = devm_request_irq(&pdev->dev, atxdmac->irq, at_xdmac_interrupt, 0,
> > > +			       "at_xdmac", atxdmac);
> > and this can invoke your irq handler and you may not be ready. Second this
> > cause races with tasklet, so usualy recommednation is to use regular
> > request_irq()
> 
> You mean it is specific to devm variant, isn't it?
Yes.

> > > +	dma_cap_set(DMA_CYCLIC, atxdmac->dma.cap_mask);
> > > +	dma_cap_set(DMA_MEMCPY, atxdmac->dma.cap_mask);
> > > +	dma_cap_set(DMA_SLAVE, atxdmac->dma.cap_mask);
> > > +	atxdmac->dma.dev				= &pdev->dev;
> > > +	atxdmac->dma.device_alloc_chan_resources	= at_xdmac_alloc_chan_resources;
> > > +	atxdmac->dma.device_free_chan_resources		= at_xdmac_free_chan_resources;
> > > +	atxdmac->dma.device_tx_status			= at_xdmac_tx_status;
> > > +	atxdmac->dma.device_issue_pending		= at_xdmac_issue_pending;
> > > +	atxdmac->dma.device_prep_dma_cyclic		= at_xdmac_prep_dma_cyclic;
> > > +	atxdmac->dma.device_prep_dma_memcpy		= at_xdmac_prep_dma_memcpy;
> > > +	atxdmac->dma.device_prep_slave_sg		= at_xdmac_prep_slave_sg;
> > > +	atxdmac->dma.device_control			= at_xdmac_control;
> > > +	atxdmac->dma.chancnt				= nr_channels;
> > no caps, pls do implement that as you have cyclic dma too
> 
> Sorry I don't understand what you mean here.
pls implement .device_slave_caps

> > > +static int at_xdmac_remove(struct platform_device *pdev)
> > > +{
> > > +	struct at_xdmac	*atxdmac = (struct at_xdmac*)platform_get_drvdata(pdev);
> > > +	int		i;
> > > +
> > > +	at_xdmac_off(atxdmac);
> > > +	of_dma_controller_free(pdev->dev.of_node);
> > > +	dma_async_device_unregister(&atxdmac->dma);
> > > +	clk_disable_unprepare(atxdmac->clk);
> > > +
> > > +	synchronize_irq(atxdmac->irq);
> > > +
> > > +	for (i = 0; i < atxdmac->dma.chancnt; i++) {
> > > +		struct at_xdmac_chan *atchan = &atxdmac->chan[i];
> > > +
> > > +		tasklet_kill(&atchan->tasklet);
> > > +		at_xdmac_free_chan_resources(&atchan->chan);
> > > +	}
> > and you can still get irq!
> 
> Why? I thought deactivating irq on device side and calling synchronize
> irq will do the stuff.
what if we have buggy hardware or spurious interrupt!

> 
> > > +
> > > +	return 0;
> > > +}
> > > +
> > > +static const struct dev_pm_ops atmel_xdmac_dev_pm_ops = {
> > > +	.prepare	= atmel_xdmac_prepare,
> > > +	.suspend_noirq	= atmel_xdmac_suspend_noirq,
> > > +	.resume_noirq	= atmel_xdmac_resume_noirq,
> > > +};
> > no ifdef CONFIG_PM?
> 
> I'll add it.
> 
> > You might want to make them late_suspend and early_resume to allow
> > clients suspending first
> 
> Ok, as I told before I was wondering which variant to use. For my
> knowledge, it will be the case also with noirq, isn't it? It is called
> after suspend_late.
Yes, thats why it makes sense for dmaengine drivers to use only late_suspend
and ealy_resume
Ludovic Desroches Sept. 10, 2014, 1:23 p.m. UTC | #4
Hi,

Sorry for the delay. I am about to send a new version but I have still
have some questions.

On Mon, Jul 28, 2014 at 10:17:50PM +0530, Vinod Koul wrote:
> On Mon, Jul 28, 2014 at 05:54:31PM +0200, Ludovic Desroches wrote:
> > > > +{
> > > > +	struct at_xdmac_desc	*desc = txd_to_at_desc(tx);
> > > > +	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(tx->chan);
> > > > +	dma_cookie_t		cookie;
> > > > +	unsigned long		flags;
> > > > +
> > > > +	spin_lock_irqsave(&atchan->lock, flags);
> > > > +	cookie = dma_cookie_assign(tx);
> > > > +
> > > > +	dev_vdbg(chan2dev(tx->chan), "%s: atchan 0x%p, add desc 0x%p to xfers_list\n",
> > > > +		 __func__, atchan, desc);
> > > > +	list_add_tail(&desc->xfer_node, &atchan->xfers_list);
> > > > +	if (list_is_singular(&atchan->xfers_list))
> > > > +		at_xdmac_start_xfer(atchan, desc);
> > > so you dont start if list has more than 1 descriptors, why?
> > Because I will let at_xdmac_advance_work managing the queue if not
> > empty. So the only moment when I have to start the transfer in tx_submit
> > is when I have only one transfer in the queue.
> So xfers_list is current active descriptors right and if it is always going
> to be singular why bother with a list?

xfers_list won't be always singular. I put all transfers into this list.
If it is not singular, I don't have to start this transfer now.

> > > > +static struct dma_async_tx_descriptor *
> > > > +at_xdmac_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
> > > > +			 size_t len, unsigned long flags)
> > > > +{
> > > > +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> > > > +	struct at_xdmac_desc	*first = NULL, *prev = NULL;
> > > > +	size_t			remaining_size = len, xfer_size = 0, ublen;
> > > > +	dma_addr_t		src_addr = src, dst_addr = dest;
> > > > +	u32			dwidth;
> > > > +	/*
> > > > +	 * WARNING: The channel configuration is set here since there is no
> > > > +	 * dmaengine_slave_config call in this case. Moreover we don't know the
> > > > +	 * direction, it involves we can't dynamically set the source and dest
> > > > +	 * interface so we have to use the same one. Only interface 0 allows EBI
> > > > +	 * access. Hopefully we can access DDR through both ports (at least on
> > > > +	 * SAMA5D4x), so we can use the same interface for source and dest,
> > > > +	 * that solves the fact we don't know the direction.
> > > and why do you need slave config for memcpy?
> > Maybe the comments is not clear. With slave config we have the direction
> > per to mem or mem to per. Of course when doing memcpy we don't need this
> > information. But I use this information not only to set the transfer
> > direction but also to set the interfaces to use for the source and dest
> > (these interfaces depend on the connection on the matrix).
> > So slave config was useful due to direction field  to tell which interface
> > to use as source and which one as dest.
> > I am lucky because NAND and DDR are on the same interfaces so it's not a
> > problem.
> > 
> > This comment is more a warning for myself to keep in mind this possible 
> > issue in order to not have a device with NAND and DDR on different
> > interfaces because it will be difficult to tell which interface is the
> > source and which one is the dest.
> Well in that case shouldnt NAND be treated like periphral and not memory?

Yes maybe it would be better but it is not needed at the moment.

[...]

> > > > +	 */
> > > > +	list_del(&desc->xfer_node);
> > > > +	list_splice_init(&desc->descs_list, &atchan->free_descs_list);
> > > shouldnt you move all the desc in active list in one shot ?
> > > 
> > 
> > Sorry I don't get it. What are you suggesting by one shot?
> Rather than invoking this per descriptor moving descriptors to
> free_desc_list one by one, we can move all descriptors togther.

Per descriptor? list_splice_init move all the desc in one shot. Are you
talking about the terminate all case? 

> > > > +static void at_xdmac_tasklet(unsigned long data)
> > > > +{
> > > > +	struct at_xdmac_chan	*atchan = (struct at_xdmac_chan *)data;
> > > > +	struct at_xdmac_desc	*desc;
> > > > +	u32			error_mask;
> > > > +
> > > > +	dev_dbg(chan2dev(&atchan->chan), "%s: status=0x%08lx\n",
> > > > +		 __func__, atchan->status);
> > > > +
> > > > +	error_mask = AT_XDMAC_CIS_RBEIS
> > > > +		     | AT_XDMAC_CIS_WBEIS
> > > > +		     | AT_XDMAC_CIS_ROIS;
> > > > +
> > > > +	if (at_xdmac_chan_is_cyclic(atchan)) {
> > > > +		at_xdmac_handle_cyclic(atchan);
> > > > +	} else if ((atchan->status & AT_XDMAC_CIS_LIS)
> > > > +		   || (atchan->status & error_mask)) {
> > > > +		struct dma_async_tx_descriptor  *txd;
> > > > +
> > > > +		if (atchan->status & AT_XDMAC_CIS_RBEIS)
> > > > +			dev_err(chan2dev(&atchan->chan), "read bus error!!!");
> > > > +		else if (atchan->status & AT_XDMAC_CIS_WBEIS)
> > > > +			dev_err(chan2dev(&atchan->chan), "write bus error!!!");
> > > > +		else if (atchan->status & AT_XDMAC_CIS_ROIS)
> > > > +			dev_err(chan2dev(&atchan->chan), "request overflow error!!!");
> > > > +
> > > > +		desc = list_first_entry(&atchan->xfers_list,
> > > > +					struct at_xdmac_desc,
> > > > +					xfer_node);
> > > > +		dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
> > > > +		BUG_ON(!desc->active_xfer);
> > > > +
> > > > +		txd = &desc->tx_dma_desc;
> > > > +
> > > > +		at_xdmac_terminate_xfer(atchan, desc);
> > > why terminate here? This looks wrong
> > 
> > Why? What do you have in mind? It is not obvious for me that it is the
> > wrong place.
> The descriptor has completed. The terminate has special meaning. The user
> can cancel all transactions by invoking terminate_all(). That implies to
> clean up the channel and abort the transfers. The transfer is done so what
> does it mean to terminate.

So maybe I have to rename this function because I need to remove the
transfer from the transfer list and put the transfer descriptors in the
free descriptor list once transfer is done.


Thanks

Ludovic
Vinod Koul Sept. 11, 2014, 5:14 a.m. UTC | #5
On Wed, Sep 10, 2014 at 03:23:53PM +0200, Ludovic Desroches wrote:
> Hi,
> 
> Sorry for the delay. I am about to send a new version but I have still
> have some questions.
> 
> On Mon, Jul 28, 2014 at 10:17:50PM +0530, Vinod Koul wrote:
> > On Mon, Jul 28, 2014 at 05:54:31PM +0200, Ludovic Desroches wrote:
> > > > > +{
> > > > > +	struct at_xdmac_desc	*desc = txd_to_at_desc(tx);
> > > > > +	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(tx->chan);
> > > > > +	dma_cookie_t		cookie;
> > > > > +	unsigned long		flags;
> > > > > +
> > > > > +	spin_lock_irqsave(&atchan->lock, flags);
> > > > > +	cookie = dma_cookie_assign(tx);
> > > > > +
> > > > > +	dev_vdbg(chan2dev(tx->chan), "%s: atchan 0x%p, add desc 0x%p to xfers_list\n",
> > > > > +		 __func__, atchan, desc);
> > > > > +	list_add_tail(&desc->xfer_node, &atchan->xfers_list);
> > > > > +	if (list_is_singular(&atchan->xfers_list))
> > > > > +		at_xdmac_start_xfer(atchan, desc);
> > > > so you dont start if list has more than 1 descriptors, why?
> > > Because I will let at_xdmac_advance_work managing the queue if not
> > > empty. So the only moment when I have to start the transfer in tx_submit
> > > is when I have only one transfer in the queue.
> > So xfers_list is current active descriptors right and if it is always going
> > to be singular why bother with a list?
> 
> xfers_list won't be always singular. I put all transfers into this list.
> If it is not singular, I don't have to start this transfer now.
Okay i think i got it now, you want to start only when its sigular and if
list is larger then at_xdmac_advance_work() would do so

> > > > > +static struct dma_async_tx_descriptor *
> > > > > +at_xdmac_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
> > > > > +			 size_t len, unsigned long flags)
> > > > > +{
> > > > > +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> > > > > +	struct at_xdmac_desc	*first = NULL, *prev = NULL;
> > > > > +	size_t			remaining_size = len, xfer_size = 0, ublen;
> > > > > +	dma_addr_t		src_addr = src, dst_addr = dest;
> > > > > +	u32			dwidth;
> > > > > +	/*
> > > > > +	 * WARNING: The channel configuration is set here since there is no
> > > > > +	 * dmaengine_slave_config call in this case. Moreover we don't know the
> > > > > +	 * direction, it involves we can't dynamically set the source and dest
> > > > > +	 * interface so we have to use the same one. Only interface 0 allows EBI
> > > > > +	 * access. Hopefully we can access DDR through both ports (at least on
> > > > > +	 * SAMA5D4x), so we can use the same interface for source and dest,
> > > > > +	 * that solves the fact we don't know the direction.
> > > > and why do you need slave config for memcpy?
> > > Maybe the comments is not clear. With slave config we have the direction
> > > per to mem or mem to per. Of course when doing memcpy we don't need this
> > > information. But I use this information not only to set the transfer
> > > direction but also to set the interfaces to use for the source and dest
> > > (these interfaces depend on the connection on the matrix).
> > > So slave config was useful due to direction field  to tell which interface
> > > to use as source and which one as dest.
> > > I am lucky because NAND and DDR are on the same interfaces so it's not a
> > > problem.
> > > 
> > > This comment is more a warning for myself to keep in mind this possible 
> > > issue in order to not have a device with NAND and DDR on different
> > > interfaces because it will be difficult to tell which interface is the
> > > source and which one is the dest.
> > Well in that case shouldnt NAND be treated like periphral and not memory?
> 
> Yes maybe it would be better but it is not needed at the moment.
> 
> [...]
> 
> > > > > +	 */
> > > > > +	list_del(&desc->xfer_node);
> > > > > +	list_splice_init(&desc->descs_list, &atchan->free_descs_list);
> > > > shouldnt you move all the desc in active list in one shot ?
> > > > 
> > > 
> > > Sorry I don't get it. What are you suggesting by one shot?
> > Rather than invoking this per descriptor moving descriptors to
> > free_desc_list one by one, we can move all descriptors togther.
> 
> Per descriptor? list_splice_init move all the desc in one shot. Are you
> talking about the terminate all case? 
In this case you invoke the function for a descriptor which you delete from
its list and then join it to free_descs_list.

> > > > > +static void at_xdmac_tasklet(unsigned long data)
> > > > > +{
> > > > > +	struct at_xdmac_chan	*atchan = (struct at_xdmac_chan *)data;
> > > > > +	struct at_xdmac_desc	*desc;
> > > > > +	u32			error_mask;
> > > > > +
> > > > > +	dev_dbg(chan2dev(&atchan->chan), "%s: status=0x%08lx\n",
> > > > > +		 __func__, atchan->status);
> > > > > +
> > > > > +	error_mask = AT_XDMAC_CIS_RBEIS
> > > > > +		     | AT_XDMAC_CIS_WBEIS
> > > > > +		     | AT_XDMAC_CIS_ROIS;
> > > > > +
> > > > > +	if (at_xdmac_chan_is_cyclic(atchan)) {
> > > > > +		at_xdmac_handle_cyclic(atchan);
> > > > > +	} else if ((atchan->status & AT_XDMAC_CIS_LIS)
> > > > > +		   || (atchan->status & error_mask)) {
> > > > > +		struct dma_async_tx_descriptor  *txd;
> > > > > +
> > > > > +		if (atchan->status & AT_XDMAC_CIS_RBEIS)
> > > > > +			dev_err(chan2dev(&atchan->chan), "read bus error!!!");
> > > > > +		else if (atchan->status & AT_XDMAC_CIS_WBEIS)
> > > > > +			dev_err(chan2dev(&atchan->chan), "write bus error!!!");
> > > > > +		else if (atchan->status & AT_XDMAC_CIS_ROIS)
> > > > > +			dev_err(chan2dev(&atchan->chan), "request overflow error!!!");
> > > > > +
> > > > > +		desc = list_first_entry(&atchan->xfers_list,
> > > > > +					struct at_xdmac_desc,
> > > > > +					xfer_node);
> > > > > +		dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
> > > > > +		BUG_ON(!desc->active_xfer);
> > > > > +
> > > > > +		txd = &desc->tx_dma_desc;
> > > > > +
> > > > > +		at_xdmac_terminate_xfer(atchan, desc);
> > > > why terminate here? This looks wrong
> > > 
> > > Why? What do you have in mind? It is not obvious for me that it is the
> > > wrong place.
> > The descriptor has completed. The terminate has special meaning. The user
> > can cancel all transactions by invoking terminate_all(). That implies to
> > clean up the channel and abort the transfers. The transfer is done so what
> > does it mean to terminate.
> 
> So maybe I have to rename this function because I need to remove the
> transfer from the transfer list and put the transfer descriptors in the
> free descriptor list once transfer is done.
ok
Ludovic Desroches Sept. 11, 2014, 6:29 a.m. UTC | #6
On Thu, Sep 11, 2014 at 10:44:52AM +0530, Vinod Koul wrote:
> On Wed, Sep 10, 2014 at 03:23:53PM +0200, Ludovic Desroches wrote:
> > Hi,
> > 
> > Sorry for the delay. I am about to send a new version but I have still
> > have some questions.
> > 
> > On Mon, Jul 28, 2014 at 10:17:50PM +0530, Vinod Koul wrote:
> > > On Mon, Jul 28, 2014 at 05:54:31PM +0200, Ludovic Desroches wrote:
> > > > > > +{
> > > > > > +	struct at_xdmac_desc	*desc = txd_to_at_desc(tx);
> > > > > > +	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(tx->chan);
> > > > > > +	dma_cookie_t		cookie;
> > > > > > +	unsigned long		flags;
> > > > > > +
> > > > > > +	spin_lock_irqsave(&atchan->lock, flags);
> > > > > > +	cookie = dma_cookie_assign(tx);
> > > > > > +
> > > > > > +	dev_vdbg(chan2dev(tx->chan), "%s: atchan 0x%p, add desc 0x%p to xfers_list\n",
> > > > > > +		 __func__, atchan, desc);
> > > > > > +	list_add_tail(&desc->xfer_node, &atchan->xfers_list);
> > > > > > +	if (list_is_singular(&atchan->xfers_list))
> > > > > > +		at_xdmac_start_xfer(atchan, desc);
> > > > > so you dont start if list has more than 1 descriptors, why?
> > > > Because I will let at_xdmac_advance_work managing the queue if not
> > > > empty. So the only moment when I have to start the transfer in tx_submit
> > > > is when I have only one transfer in the queue.
> > > So xfers_list is current active descriptors right and if it is always going
> > > to be singular why bother with a list?
> > 
> > xfers_list won't be always singular. I put all transfers into this list.
> > If it is not singular, I don't have to start this transfer now.
> Okay i think i got it now, you want to start only when its sigular and if
> list is larger then at_xdmac_advance_work() would do so

Yes that's it.

> 
> > > > > > +static struct dma_async_tx_descriptor *
> > > > > > +at_xdmac_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
> > > > > > +			 size_t len, unsigned long flags)
> > > > > > +{
> > > > > > +	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
> > > > > > +	struct at_xdmac_desc	*first = NULL, *prev = NULL;
> > > > > > +	size_t			remaining_size = len, xfer_size = 0, ublen;
> > > > > > +	dma_addr_t		src_addr = src, dst_addr = dest;
> > > > > > +	u32			dwidth;
> > > > > > +	/*
> > > > > > +	 * WARNING: The channel configuration is set here since there is no
> > > > > > +	 * dmaengine_slave_config call in this case. Moreover we don't know the
> > > > > > +	 * direction, it involves we can't dynamically set the source and dest
> > > > > > +	 * interface so we have to use the same one. Only interface 0 allows EBI
> > > > > > +	 * access. Hopefully we can access DDR through both ports (at least on
> > > > > > +	 * SAMA5D4x), so we can use the same interface for source and dest,
> > > > > > +	 * that solves the fact we don't know the direction.
> > > > > and why do you need slave config for memcpy?
> > > > Maybe the comments is not clear. With slave config we have the direction
> > > > per to mem or mem to per. Of course when doing memcpy we don't need this
> > > > information. But I use this information not only to set the transfer
> > > > direction but also to set the interfaces to use for the source and dest
> > > > (these interfaces depend on the connection on the matrix).
> > > > So slave config was useful due to direction field  to tell which interface
> > > > to use as source and which one as dest.
> > > > I am lucky because NAND and DDR are on the same interfaces so it's not a
> > > > problem.
> > > > 
> > > > This comment is more a warning for myself to keep in mind this possible 
> > > > issue in order to not have a device with NAND and DDR on different
> > > > interfaces because it will be difficult to tell which interface is the
> > > > source and which one is the dest.
> > > Well in that case shouldnt NAND be treated like periphral and not memory?
> > 
> > Yes maybe it would be better but it is not needed at the moment.
> > 
> > [...]
> > 
> > > > > > +	 */
> > > > > > +	list_del(&desc->xfer_node);
> > > > > > +	list_splice_init(&desc->descs_list, &atchan->free_descs_list);
> > > > > shouldnt you move all the desc in active list in one shot ?
> > > > > 
> > > > 
> > > > Sorry I don't get it. What are you suggesting by one shot?
> > > Rather than invoking this per descriptor moving descriptors to
> > > free_desc_list one by one, we can move all descriptors togther.
> > 
> > Per descriptor? list_splice_init move all the desc in one shot. Are you
> > talking about the terminate all case? 
> In this case you invoke the function for a descriptor which you delete from
> its list and then join it to free_descs_list.
> 

I remove the transfer from the transfers list then I put the descriptors of this
transfer in the free descriptors list. I am playing with two different
lists.

> > > > > > +static void at_xdmac_tasklet(unsigned long data)
> > > > > > +{
> > > > > > +	struct at_xdmac_chan	*atchan = (struct at_xdmac_chan *)data;
> > > > > > +	struct at_xdmac_desc	*desc;
> > > > > > +	u32			error_mask;
> > > > > > +
> > > > > > +	dev_dbg(chan2dev(&atchan->chan), "%s: status=0x%08lx\n",
> > > > > > +		 __func__, atchan->status);
> > > > > > +
> > > > > > +	error_mask = AT_XDMAC_CIS_RBEIS
> > > > > > +		     | AT_XDMAC_CIS_WBEIS
> > > > > > +		     | AT_XDMAC_CIS_ROIS;
> > > > > > +
> > > > > > +	if (at_xdmac_chan_is_cyclic(atchan)) {
> > > > > > +		at_xdmac_handle_cyclic(atchan);
> > > > > > +	} else if ((atchan->status & AT_XDMAC_CIS_LIS)
> > > > > > +		   || (atchan->status & error_mask)) {
> > > > > > +		struct dma_async_tx_descriptor  *txd;
> > > > > > +
> > > > > > +		if (atchan->status & AT_XDMAC_CIS_RBEIS)
> > > > > > +			dev_err(chan2dev(&atchan->chan), "read bus error!!!");
> > > > > > +		else if (atchan->status & AT_XDMAC_CIS_WBEIS)
> > > > > > +			dev_err(chan2dev(&atchan->chan), "write bus error!!!");
> > > > > > +		else if (atchan->status & AT_XDMAC_CIS_ROIS)
> > > > > > +			dev_err(chan2dev(&atchan->chan), "request overflow error!!!");
> > > > > > +
> > > > > > +		desc = list_first_entry(&atchan->xfers_list,
> > > > > > +					struct at_xdmac_desc,
> > > > > > +					xfer_node);
> > > > > > +		dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
> > > > > > +		BUG_ON(!desc->active_xfer);
> > > > > > +
> > > > > > +		txd = &desc->tx_dma_desc;
> > > > > > +
> > > > > > +		at_xdmac_terminate_xfer(atchan, desc);
> > > > > why terminate here? This looks wrong
> > > > 
> > > > Why? What do you have in mind? It is not obvious for me that it is the
> > > > wrong place.
> > > The descriptor has completed. The terminate has special meaning. The user
> > > can cancel all transactions by invoking terminate_all(). That implies to
> > > clean up the channel and abort the transfers. The transfer is done so what
> > > does it mean to terminate.
> > 
> > So maybe I have to rename this function because I need to remove the
> > transfer from the transfer list and put the transfer descriptors in the
> > free descriptor list once transfer is done.
> ok
> 
> -- 
> ~Vinod
> 
> --
> To unsubscribe from this list: send the line "unsubscribe dmaengine" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/dma/Kconfig b/drivers/dma/Kconfig
index 1eca7b9..068360c 100644
--- a/drivers/dma/Kconfig
+++ b/drivers/dma/Kconfig
@@ -89,6 +89,13 @@  config AT_HDMAC
 	help
 	  Support the Atmel AHB DMA controller.
 
+config AT_XDMAC
+	tristate "Atmel XDMA support"
+	depends on ARCH_AT91
+	select DMA_ENGINE
+	help
+	  Support the Atmel XDMA controller.
+
 config FSL_DMA
 	tristate "Freescale Elo series DMA support"
 	depends on FSL_SOC
diff --git a/drivers/dma/Makefile b/drivers/dma/Makefile
index c779e1e..2048eb8 100644
--- a/drivers/dma/Makefile
+++ b/drivers/dma/Makefile
@@ -17,6 +17,7 @@  obj-$(CONFIG_PPC_BESTCOMM) += bestcomm/
 obj-$(CONFIG_MV_XOR) += mv_xor.o
 obj-$(CONFIG_DW_DMAC_CORE) += dw/
 obj-$(CONFIG_AT_HDMAC) += at_hdmac.o
+obj-$(CONFIG_AT_XDMAC) += at_xdmac.o
 obj-$(CONFIG_MX3_IPU) += ipu/
 obj-$(CONFIG_TXX9_DMAC) += txx9dmac.o
 obj-$(CONFIG_SH_DMAE_BASE) += sh/
diff --git a/drivers/dma/at_xdmac.c b/drivers/dma/at_xdmac.c
new file mode 100644
index 0000000..71d371d
--- /dev/null
+++ b/drivers/dma/at_xdmac.c
@@ -0,0 +1,1181 @@ 
+/*
+ * Driver for the Atmel Extensible DMA Controller (aka XDMAC on AT91 systems)
+ *
+ * Copyright (C) 2014 Atmel Corporation
+ *
+ * Author: Ludovic Desroches <ludovic.desroches@atmel.com>
+ *
+ * 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.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include <dt-bindings/dma/at91.h>
+#include <linux/clk.h>
+#include <linux/dmaengine.h>
+#include <linux/dmapool.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/of_dma.h>
+#include <linux/of_platform.h>
+#include <linux/platform_device.h>
+
+#include "dmaengine.h"
+#include "at_xdmac.h"
+
+static unsigned int init_nr_desc_per_channel = 64;
+module_param(init_nr_desc_per_channel, uint, 0644);
+MODULE_PARM_DESC(init_nr_desc_per_channel,
+		 "initial descriptors per channel (default: 64)");
+
+
+static bool at_xdmac_chan_is_enabled(struct at_xdmac_chan *atchan)
+{
+	return at_xdmac_chan_read(atchan, AT_XDMAC_GS) & atchan->mask;
+}
+
+static void at_xdmac_off(struct at_xdmac *atxdmac)
+{
+	at_xdmac_write(atxdmac, AT_XDMAC_GD, -1L);
+
+	/* Wait that all chans are disabled. */
+	while (at_xdmac_read(atxdmac, AT_XDMAC_GS))
+		cpu_relax();
+
+	at_xdmac_write(atxdmac, AT_XDMAC_GID, -1L);
+}
+
+/* Call with lock hold. */
+static void at_xdmac_start_xfer(struct at_xdmac_chan *atchan,
+				struct at_xdmac_desc *first)
+{
+	struct at_xdmac	*atxdmac = to_at_xdmac(atchan->chan.device);
+	u32		reg;
+
+	dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, first);
+
+	if (at_xdmac_chan_is_enabled(atchan)) {
+		dev_err(chan2dev(&atchan->chan),
+			"BUG: Attempted to start a non-idle channel\n");
+		return;
+	}
+
+	/* Set transfer as active to not try to start it again. */
+	first->active_xfer = true;
+
+	/* Tell xdmac where to get the first descriptor. */
+	reg = AT_XDMAC_CNDA_NDA(first->tx_dma_desc.phys)
+	      | AT_XDMAC_CNDA_NDAIF(atchan->memif);
+	at_xdmac_chan_write(atchan, AT_XDMAC_CNDA, reg);
+
+	/*
+	 * When doing memory to memory transfer we need to use the next
+	 * descriptor view 2 since some fields of the configuration register
+	 * depend on transfer size and src/dest addresses.
+	 */
+	if (atchan->cfg & AT_XDMAC_CC_TYPE_PER_TRAN) {
+		reg = AT_XDMAC_CNDC_NDVIEW_NDV1;
+		at_xdmac_chan_write(atchan, AT_XDMAC_CC, atchan->cfg);
+	} else
+		reg = AT_XDMAC_CNDC_NDVIEW_NDV2;
+
+	reg |= AT_XDMAC_CNDC_NDDUP
+	       | AT_XDMAC_CNDC_NDSUP
+	       | AT_XDMAC_CNDC_NDE;
+	at_xdmac_chan_write(atchan, AT_XDMAC_CNDC, reg);
+
+	dev_vdbg(chan2dev(&atchan->chan),
+		 "%s: XDMAC_CC=0x%08x XDMAC_CNDA=0x%08x, XDMAC_CNDC=0x%08x, "
+		 "XDMAC_CSA=0x%08x, XDMAC_CDA=0x%08x, XDMAC_CUBC=0x%08x\n",
+		 __func__, at_xdmac_chan_read(atchan, AT_XDMAC_CC),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CNDA),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CNDC),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CSA),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CDA),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CUBC));
+
+	/*
+	 * There is no end of list when doing cyclic dma, we need to get
+	 * an interrupt after each periods.
+	 */
+	if (at_xdmac_chan_is_cyclic(atchan))
+		at_xdmac_chan_write(atchan, AT_XDMAC_CIE, AT_XDMAC_CIE_BIE);
+	else
+		at_xdmac_chan_write(atchan, AT_XDMAC_CIE, AT_XDMAC_CIE_LIE);
+	at_xdmac_write(atxdmac, AT_XDMAC_GIE, atchan->mask);
+	dev_vdbg(chan2dev(&atchan->chan),
+		 "%s: enable channel (0x%08x)\n", __func__, atchan->mask);
+	at_xdmac_write(atxdmac, AT_XDMAC_GE, atchan->mask);
+
+	dev_vdbg(chan2dev(&atchan->chan),
+		 "%s: XDMAC_CC=0x%08x XDMAC_CNDA=0x%08x, XDMAC_CNDC=0x%08x, "
+		 "XDMAC_CSA=0x%08x, XDMAC_CDA=0x%08x, XDMAC_CUBC=0x%08x\n",
+		 __func__, at_xdmac_chan_read(atchan, AT_XDMAC_CC),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CNDA),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CNDC),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CSA),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CDA),
+		 at_xdmac_chan_read(atchan, AT_XDMAC_CUBC));
+
+}
+
+static dma_cookie_t at_xdmac_tx_submit(struct dma_async_tx_descriptor *tx)
+{
+	struct at_xdmac_desc	*desc = txd_to_at_desc(tx);
+	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(tx->chan);
+	dma_cookie_t		cookie;
+	unsigned long		flags;
+
+	spin_lock_irqsave(&atchan->lock, flags);
+	cookie = dma_cookie_assign(tx);
+
+	dev_vdbg(chan2dev(tx->chan), "%s: atchan 0x%p, add desc 0x%p to xfers_list\n",
+		 __func__, atchan, desc);
+	list_add_tail(&desc->xfer_node, &atchan->xfers_list);
+	if (list_is_singular(&atchan->xfers_list))
+		at_xdmac_start_xfer(atchan, desc);
+
+	spin_unlock_irqrestore(&atchan->lock, flags);
+	return cookie;
+}
+
+static struct at_xdmac_desc *at_xdmac_alloc_desc(struct dma_chan *chan,
+						 gfp_t gfp_flags)
+{
+	struct at_xdmac_desc	*desc;
+	struct at_xdmac		*atxdmac = to_at_xdmac(chan->device);
+	dma_addr_t		phys;
+
+	desc = dma_pool_alloc(atxdmac->at_xdmac_desc_pool, gfp_flags, &phys);
+	if (desc) {
+		memset(desc, 0, sizeof(*desc));
+		INIT_LIST_HEAD(&desc->descs_list);
+		dma_async_tx_descriptor_init(&desc->tx_dma_desc, chan);
+		desc->tx_dma_desc.tx_submit = at_xdmac_tx_submit;
+		desc->tx_dma_desc.phys = phys;
+	}
+
+	return desc;
+}
+
+/* Call must be protected by lock. */
+static struct at_xdmac_desc *at_xdmac_get_desc(struct at_xdmac_chan *atchan)
+{
+	struct at_xdmac_desc *desc;
+
+	if (list_empty(&atchan->free_descs_list)) {
+		desc = at_xdmac_alloc_desc(&atchan->chan, GFP_ATOMIC);
+	} else {
+		desc = list_first_entry(&atchan->free_descs_list,
+					struct at_xdmac_desc, desc_node);
+		list_del(&desc->desc_node);
+	}
+
+	return desc;
+}
+
+static bool at_xdmac_filter(struct dma_chan *chan, void *slave)
+{
+	return chan->device->dev == slave;
+}
+
+static struct dma_chan *at_xdmac_xlate(struct of_phandle_args *dma_spec,
+				       struct of_dma *of_dma)
+{
+	struct at_xdmac		*atxdmac = of_dma->of_dma_data;
+	struct at_xdmac_chan	*atchan;
+	struct dma_chan		*chan;
+	struct device		*dev = atxdmac->dma.dev;
+	dma_cap_mask_t 		mask;
+
+	if (dma_spec->args_count != 2) {
+		dev_err(dev, "dma phandler args: bad number of args\n");
+		return NULL;
+	}
+
+	dma_cap_zero(mask);
+	dma_cap_set(DMA_SLAVE, mask);
+	chan = dma_request_channel(mask, at_xdmac_filter, dev);
+	if (!chan) {
+		dev_err(dev, "can't get a dma channel\n");
+		return NULL;
+	}
+
+	atchan = to_at_xdmac_chan(chan);
+	atchan->memif = AT91_XDMAC_DT_GET_MEM_IF(dma_spec->args[0]);
+	atchan->perif = AT91_XDMAC_DT_GET_PER_IF(dma_spec->args[0]);
+	atchan->perid = AT91_XDMAC_DT_GET_PERID(dma_spec->args[1]);
+	dev_info(dev, "chan dt cfg: memif=%u perif=%u perid=%u\n",
+		 atchan->memif, atchan->perif, atchan->perid);
+
+	return chan;
+}
+
+static int at_xdmac_set_slave_config(struct dma_chan *chan,
+				      struct dma_slave_config *sconfig)
+{
+	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
+
+	atchan->cfg = AT91_XDMAC_DT_PERID(atchan->perid)
+		      |	AT_XDMAC_CC_SWREQ_HWR_CONNECTED
+		      | AT_XDMAC_CC_MBSIZE_SIXTEEN
+		      | AT_XDMAC_CC_TYPE_PER_TRAN;
+
+	if (sconfig->direction == DMA_DEV_TO_MEM) {
+		atchan->cfg |= AT_XDMAC_CC_DAM_INCREMENTED_AM
+			       | AT_XDMAC_CC_SAM_FIXED_AM
+			       | AT_XDMAC_CC_DIF(atchan->memif)
+			       | AT_XDMAC_CC_SIF(atchan->perif)
+			       | AT_XDMAC_CC_DSYNC_PER2MEM;
+		atchan->dwidth = ffs(sconfig->src_addr_width) - 1;
+		atchan->cfg |= AT_XDMAC_CC_DWIDTH(atchan->dwidth);
+		atchan->cfg |= at_xdmac_csize(sconfig->src_maxburst);
+	} else if (sconfig->direction == DMA_MEM_TO_DEV) {
+		atchan->cfg |= AT_XDMAC_CC_DAM_FIXED_AM
+			       | AT_XDMAC_CC_SAM_INCREMENTED_AM
+			       | AT_XDMAC_CC_DIF(atchan->perif)
+			       | AT_XDMAC_CC_SIF(atchan->memif)
+			       | AT_XDMAC_CC_DSYNC_MEM2PER;
+		atchan->dwidth = ffs(sconfig->dst_addr_width) - 1;
+		atchan->cfg |= AT_XDMAC_CC_DWIDTH(atchan->dwidth);
+		atchan->cfg |= at_xdmac_csize(sconfig->dst_maxburst);
+	} else
+		return -EINVAL;
+
+	/*
+	 * Src address and dest addr are needed to configure the link list
+	 * descriptor so keep the slave configuration.
+	 */
+	memcpy(&atchan->dma_sconfig, sconfig, sizeof(struct dma_slave_config));
+
+	dev_dbg(chan2dev(chan), "%s: atchan->cfg=0x%08x\n", __func__, atchan->cfg);
+
+	return 0;
+}
+
+static struct dma_async_tx_descriptor *
+at_xdmac_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,
+		       unsigned int sg_len, enum dma_transfer_direction direction,
+		       unsigned long flags, void *context)
+{
+	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
+	struct dma_slave_config	*sconfig = &atchan->dma_sconfig;
+	struct at_xdmac_desc	*first = NULL, *prev = NULL;
+	struct scatterlist	*sg;
+	int 			i;
+
+	if (!sgl)
+		return NULL;
+
+	if (!is_slave_direction(direction)) {
+		dev_err(chan2dev(chan), "invalid DMA direction\n");
+		return NULL;
+	}
+
+	dev_dbg(chan2dev(chan), "%s: sg_len=%d, dir=%s, flags=0x%lx\n",
+		 __func__, sg_len,
+		 direction == DMA_MEM_TO_DEV ? "to device" : "from device",
+		 flags);
+
+	/* Protect dma_sconfig field that can be modified by set_slave_conf. */
+	spin_lock(&atchan->lock);
+
+	/* Prepare descriptors. */
+	for_each_sg(sgl, sg, sg_len, i) {
+		struct at_xdmac_desc	*desc = NULL;
+		u32			len, mem;
+
+		len = sg_dma_len(sg);
+		mem = sg_dma_address(sg);
+		if (unlikely(!len)) {
+			dev_err(chan2dev(chan), "sg data length is zero\n");
+			spin_unlock(&atchan->lock);
+			return NULL;
+		}
+		dev_dbg(chan2dev(chan), "%s: * sg%d len=%u, mem=0x%08x\n",
+			 __func__, i, len, mem);
+
+		desc = at_xdmac_get_desc(atchan);
+		if (!desc) {
+			dev_err(chan2dev(chan), "can't get descriptor\n");
+			if (first)
+				list_splice_init(&first->descs_list, &atchan->free_descs_list);
+			spin_unlock(&atchan->lock);
+			return NULL;
+		}
+
+		/* Linked list descriptor setup. */
+		if (direction == DMA_DEV_TO_MEM) {
+			desc->lld.mbr_sa = sconfig->src_addr;
+			desc->lld.mbr_da = mem;
+		} else {
+			desc->lld.mbr_sa = mem;
+			desc->lld.mbr_da = sconfig->dst_addr;
+		}
+		desc->lld.mbr_ubc = AT_XDMAC_MBR_UBC_NDV1		/* next descriptor view */
+			| AT_XDMAC_MBR_UBC_NDEN				/* next descriptor dst parameter update */
+			| AT_XDMAC_MBR_UBC_NSEN				/* next descriptor src parameter update */
+			| (i == sg_len - 1 ? 0 : AT_XDMAC_MBR_UBC_NDE)	/* descriptor fetch */
+			| len / (1 << atchan->dwidth);			/* microblock length */
+		dev_dbg(chan2dev(chan),
+			 "%s: lld: mbr_sa=0x%08x, mbr_da=0x%08x, mbr_ubc=0x%08x\n",
+			 __func__, desc->lld.mbr_sa, desc->lld.mbr_da, desc->lld.mbr_ubc);
+
+		/* Chain lld. */
+		if (prev) {
+			prev->lld.mbr_nda = desc->tx_dma_desc.phys;
+			dev_dbg(chan2dev(chan),
+				 "%s: chain lld: prev=0x%p, mbr_nda=0x%08x\n",
+				 __func__, prev, prev->lld.mbr_nda);
+		}
+
+		prev = desc;
+		if (!first)
+			first = desc;
+
+		dev_dbg(chan2dev(chan), "%s: add desc 0x%p to descs_list 0x%p\n",
+			 __func__, desc, first);
+		list_add_tail(&desc->desc_node, &first->descs_list);
+	}
+
+	spin_unlock(&atchan->lock);
+
+	first->tx_dma_desc.cookie = -EBUSY;
+	first->tx_dma_desc.flags = flags;
+	first->xfer_size = sg_len;
+
+	return &first->tx_dma_desc;
+}
+
+static struct dma_async_tx_descriptor *
+at_xdmac_prep_dma_cyclic(struct dma_chan *chan, dma_addr_t buf_addr,
+			 size_t buf_len, size_t period_len,
+			 enum dma_transfer_direction direction,
+			 unsigned long flags, void *context)
+{
+	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
+	struct dma_slave_config	*sconfig = &atchan->dma_sconfig;
+	struct at_xdmac_desc	*first = NULL, *prev = NULL;
+	unsigned int		periods = buf_len / period_len;
+	unsigned long		lock_flags;
+	int 			i;
+
+	dev_dbg(chan2dev(chan), "%s: buf_addr=0x%08x, buf_len=%d, period_len=%d, "
+		"dir=%s, flags=0x%lx\n",
+		__func__, buf_addr, buf_len, period_len,
+		direction == DMA_MEM_TO_DEV ? "mem2per" : "per2mem", flags);
+
+	if (!is_slave_direction(direction)) {
+		dev_err(chan2dev(chan), "invalid DMA direction\n");
+		return NULL;
+	}
+
+	if (test_and_set_bit(AT_XDMAC_CHAN_IS_CYCLIC, &atchan->status)) {
+		dev_err(chan2dev(chan), "channel currently used\n");
+		return NULL;
+	}
+
+	for (i = 0; i < periods; i++) {
+		struct at_xdmac_desc	*desc = NULL;
+
+		spin_lock_irqsave(&atchan->lock, lock_flags);
+		desc = at_xdmac_get_desc(atchan);
+		spin_unlock_irqrestore(&atchan->lock, lock_flags);
+		if (!desc) {
+			dev_err(chan2dev(chan), "can't get descriptor\n");
+			if (first)
+				list_splice_init(&first->descs_list, &atchan->free_descs_list);
+			return NULL;
+		}
+		dev_dbg(chan2dev(chan),
+			"%s: desc=0x%p, tx_dma_desc.phys=0x%08x\n",
+			__func__, desc, desc->tx_dma_desc.phys);
+
+		if (direction == DMA_DEV_TO_MEM) {
+			desc->lld.mbr_sa = sconfig->src_addr;
+			desc->lld.mbr_da = buf_addr + i * period_len;
+		} else {
+			desc->lld.mbr_sa = buf_addr + i * period_len;
+			desc->lld.mbr_da = sconfig->dst_addr;
+		};
+		desc->lld.mbr_ubc = AT_XDMAC_MBR_UBC_NDV1
+			| AT_XDMAC_MBR_UBC_NDEN
+			| AT_XDMAC_MBR_UBC_NSEN
+			| AT_XDMAC_MBR_UBC_NDE
+			| period_len / (1 << atchan->dwidth);
+
+		dev_dbg(chan2dev(chan),
+			 "%s: lld: mbr_sa=0x%08x, mbr_da=0x%08x, mbr_ubc=0x%08x\n",
+			 __func__, desc->lld.mbr_sa, desc->lld.mbr_da, desc->lld.mbr_ubc);
+
+		/* Chain lld. */
+		if (prev) {
+			prev->lld.mbr_nda = desc->tx_dma_desc.phys;
+			dev_dbg(chan2dev(chan),
+				 "%s: chain lld: prev=0x%p, mbr_nda=0x%08x\n",
+				 __func__, prev, prev->lld.mbr_nda);
+		}
+
+		prev = desc;
+		if (!first)
+			first = desc;
+
+		dev_dbg(chan2dev(chan), "%s: add desc 0x%p to descs_list 0x%p\n",
+			 __func__, desc, first);
+		list_add_tail(&desc->desc_node, &first->descs_list);
+	}
+
+	prev->lld.mbr_nda = first->tx_dma_desc.phys;
+	dev_dbg(chan2dev(chan),
+		"%s: chain lld: prev=0x%p, mbr_nda=0x%08x\n",
+		__func__, prev, prev->lld.mbr_nda);
+	first->tx_dma_desc.cookie = -EBUSY;
+	first->tx_dma_desc.flags = flags;
+	first->xfer_size = buf_len;
+
+	return &first->tx_dma_desc;
+}
+
+static struct dma_async_tx_descriptor *
+at_xdmac_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
+			 size_t len, unsigned long flags)
+{
+	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
+	struct at_xdmac_desc	*first = NULL, *prev = NULL;
+	size_t			remaining_size = len, xfer_size = 0, ublen;
+	dma_addr_t		src_addr = src, dst_addr = dest;
+	u32			dwidth;
+	/*
+	 * WARNING: The channel configuration is set here since there is no
+	 * dmaengine_slave_config call in this case. Moreover we don't know the
+	 * direction, it involves we can't dynamically set the source and dest
+	 * interface so we have to use the same one. Only interface 0 allows EBI
+	 * access. Hopefully we can access DDR through both ports (at least on
+	 * SAMA5D4x), so we can use the same interface for source and dest,
+	 * that solves the fact we don't know the direction.
+	 */
+	u32			chan_cc = AT_XDMAC_CC_DAM_INCREMENTED_AM
+					| AT_XDMAC_CC_SAM_INCREMENTED_AM
+					| AT_XDMAC_CC_DIF(0)
+					| AT_XDMAC_CC_SIF(0)
+					| AT_XDMAC_CC_MBSIZE_SIXTEEN
+					| AT_XDMAC_CC_TYPE_MEM_TRAN;
+
+	dev_dbg(chan2dev(chan), "%s: src=0x%08x, dest=0x%08x, len=%d, flags=0x%lx\n",
+		__func__, src, dest, len, flags);
+
+	if (unlikely(!len))
+		return NULL;
+
+	/* Check address alignment to select the greater data width we can use.
+	 * Some XDMAC implementations don't provide dword transfer, in this
+	 * case selecting dword has the same behavior as selecting word transfers.
+	 */
+	if (!((src_addr | dst_addr) & 7)) {
+		dwidth = AT_XDMAC_CC_DWIDTH_DWORD;
+		dev_dbg(chan2dev(chan), "%s: dwidth: double word\n", __func__);
+	} else if (!((src_addr | dst_addr)  & 3)) {
+		dwidth = AT_XDMAC_CC_DWIDTH_WORD;
+		dev_dbg(chan2dev(chan), "%s: dwidth: word\n", __func__);
+	} else if (!((src_addr | dst_addr) & 1)) {
+		dwidth = AT_XDMAC_CC_DWIDTH_HALFWORD;
+		dev_dbg(chan2dev(chan), "%s: dwidth: half word\n", __func__);
+	} else {
+		dwidth = AT_XDMAC_CC_DWIDTH_BYTE;
+		dev_dbg(chan2dev(chan), "%s: dwidth: byte\n", __func__);
+	}
+
+	atchan->cfg = chan_cc | AT_XDMAC_CC_DWIDTH(dwidth);
+
+	/* Prepare descriptors. */
+	while (remaining_size) {
+		struct at_xdmac_desc	*desc = NULL;
+
+		dev_dbg(chan2dev(chan), "%s: remaining_size=%u\n", __func__, remaining_size);
+
+		spin_lock_irqsave(&atchan->lock, flags);
+		desc = at_xdmac_get_desc(atchan);
+		spin_unlock_irqrestore(&atchan->lock, flags);
+		if (!desc) {
+			dev_err(chan2dev(chan), "can't get descriptor\n");
+			if (first)
+				list_splice_init(&first->descs_list, &atchan->free_descs_list);
+			return NULL;
+		}
+
+		/* Update src and dest addresses. */
+		src_addr += xfer_size;
+		dst_addr += xfer_size;
+
+		if (remaining_size >= AT_XDMAC_MBR_UBC_UBLEN_MAX << dwidth)
+			xfer_size = AT_XDMAC_MBR_UBC_UBLEN_MAX << dwidth;
+		else
+			xfer_size = remaining_size;
+
+		dev_dbg(chan2dev(chan), "%s: xfer_size=%u\n", __func__, xfer_size);
+
+		/* Check remaining length and change data width if needed. */
+		if (!((src_addr | dst_addr | xfer_size) & 7)) {
+			dwidth = AT_XDMAC_CC_DWIDTH_DWORD;
+			dev_dbg(chan2dev(chan), "%s: dwidth: double word\n", __func__);
+		} else if (!((src_addr | dst_addr | xfer_size)  & 3)) {
+			dwidth = AT_XDMAC_CC_DWIDTH_WORD;
+			dev_dbg(chan2dev(chan), "%s: dwidth: word\n", __func__);
+		} else if (!((src_addr | dst_addr | xfer_size) & 1)) {
+			dwidth = AT_XDMAC_CC_DWIDTH_HALFWORD;
+			dev_dbg(chan2dev(chan), "%s: dwidth: half word\n", __func__);
+		} else if ((src_addr | dst_addr | xfer_size) & 1) {
+			dwidth = AT_XDMAC_CC_DWIDTH_BYTE;
+			dev_dbg(chan2dev(chan), "%s: dwidth: byte\n", __func__);
+		}
+		chan_cc |= AT_XDMAC_CC_DWIDTH(dwidth);
+
+		ublen = xfer_size >> dwidth;
+		remaining_size -= xfer_size;
+
+		desc->lld.mbr_sa = src_addr;
+		desc->lld.mbr_da = dst_addr;
+		desc->lld.mbr_ubc = AT_XDMAC_MBR_UBC_NDV2
+			| AT_XDMAC_MBR_UBC_NDEN
+			| AT_XDMAC_MBR_UBC_NSEN
+			| (remaining_size ? 0 : AT_XDMAC_MBR_UBC_NDE)
+			| ublen;
+		desc->lld.mbr_cfg = chan_cc;
+
+		dev_dbg(chan2dev(chan),
+			 "%s: lld: mbr_sa=0x%08x, mbr_da=0x%08x, mbr_ubc=0x%08x, mbr_cfg=0x%08x\n",
+			 __func__, desc->lld.mbr_sa, desc->lld.mbr_da, desc->lld.mbr_ubc, desc->lld.mbr_cfg);
+
+		/* Chain lld. */
+		if (prev) {
+			prev->lld.mbr_nda = desc->tx_dma_desc.phys;
+			dev_dbg(chan2dev(chan),
+				 "%s: chain lld: prev=0x%p, mbr_nda=0x%08x\n",
+				 __func__, prev, prev->lld.mbr_nda);
+		}
+
+		prev = desc;
+		if (!first)
+			first = desc;
+
+		dev_dbg(chan2dev(chan), "%s: add desc 0x%p to descs_list 0x%p\n",
+			 __func__, desc, first);
+		list_add_tail(&desc->desc_node, &first->descs_list);
+	}
+
+	first->tx_dma_desc.cookie = -EBUSY;
+	first->tx_dma_desc.flags = flags;
+	first->xfer_size = len;
+
+	return &first->tx_dma_desc;
+}
+
+static enum dma_status
+at_xdmac_tx_status(struct dma_chan *chan, dma_cookie_t cookie,
+		struct dma_tx_state *txstate)
+{
+	struct at_xdmac_chan 	*atchan = to_at_xdmac_chan(chan);
+	struct at_xdmac		*atxdmac = to_at_xdmac(atchan->chan.device);
+	struct at_xdmac_desc	*desc, *_desc;
+	unsigned long		flags;
+	enum dma_status		ret;
+	int			residue;
+	u32			cur_nda;
+
+	ret = dma_cookie_status(chan, cookie, txstate);
+	if (ret == DMA_COMPLETE)
+		return ret;
+
+	spin_lock_irqsave(&atchan->lock, flags);
+
+	desc = list_first_entry(&atchan->xfers_list, struct at_xdmac_desc, xfer_node);
+
+	if (!desc->active_xfer)
+		dev_err(chan2dev(chan),
+			"something goes wrong, there is no active transfer\n");
+
+	residue = desc->xfer_size;
+
+	/* Flush FIFO. */
+	at_xdmac_write(atxdmac, AT_XDMAC_GSWF, atchan->mask);
+	while (!(at_xdmac_chan_read(atchan, AT_XDMAC_CIS) & AT_XDMAC_CIS_FIS));
+		cpu_relax();
+
+	cur_nda = at_xdmac_chan_read(atchan, AT_XDMAC_CNDA) & 0xfffffffc;
+	/*
+	 * Remove size of all microblocks already transferred and the current
+	 * one. Then add the remaining size to transfer of the current
+	 * microblock.
+	 */
+	list_for_each_entry_safe(desc, _desc, &desc->descs_list, desc_node) {
+		residue -= (desc->lld.mbr_ubc & 0xffffff) << atchan->dwidth;
+		if ((desc->lld.mbr_nda & 0xfffffffc) == cur_nda)
+			break;
+	}
+	residue += at_xdmac_chan_read(atchan, AT_XDMAC_CUBC) << atchan->dwidth;
+
+	spin_unlock_irqrestore(&atchan->lock, flags);
+
+	dma_set_residue(txstate, residue);
+
+	dev_dbg(chan2dev(chan),
+		 "%s: desc=0x%p, tx_dma_desc.phys=0x%08x, tx_status=%d, cookie=%d, residue=%d\n",
+		 __func__, desc, desc->tx_dma_desc.phys, ret, cookie, residue);
+
+	return ret;
+}
+
+static void at_xdmac_terminate_xfer(struct at_xdmac_chan *atchan,
+				    struct at_xdmac_desc *desc)
+{
+	dev_dbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
+
+	/*
+	 * It's necessary to remove the transfer before calling the callback
+	 * because some devices can call dma_engine_terminate_all causing to do
+	 * dma_cookie_complete two times on the same cookie.
+	 */
+	list_del(&desc->xfer_node);
+	list_splice_init(&desc->descs_list, &atchan->free_descs_list);
+}
+
+static void at_xdmac_advance_work(struct at_xdmac_chan *atchan)
+{
+	struct at_xdmac_desc	*desc;
+	unsigned long		flags;
+
+	spin_lock_irqsave(&atchan->lock, flags);
+
+	/*
+	 * If channel is enabled, do nothing, advance_work will be triggered
+	 * after the interruption.
+	 */
+	if (at_xdmac_chan_is_enabled(atchan)) {
+		dev_dbg(chan2dev(&atchan->chan), "%s: chan enabled\n",
+			 __func__);
+	} else if (!list_empty(&atchan->xfers_list)) {
+		desc = list_first_entry(&atchan->xfers_list,
+					struct at_xdmac_desc,
+					xfer_node);
+		dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
+		if (!desc->active_xfer)
+			at_xdmac_start_xfer(atchan, desc);
+	}
+
+	spin_unlock_irqrestore(&atchan->lock, flags);
+}
+
+static void at_xdmac_handle_cyclic(struct at_xdmac_chan *atchan)
+{
+	struct at_xdmac_desc		*desc;
+	struct dma_async_tx_descriptor	*txd;
+
+	desc = list_first_entry(&atchan->xfers_list, struct at_xdmac_desc, xfer_node);
+	txd = &desc->tx_dma_desc;
+
+	if (txd->callback && (txd->flags & DMA_PREP_INTERRUPT))
+		txd->callback(txd->callback_param);
+}
+
+static void at_xdmac_tasklet(unsigned long data)
+{
+	struct at_xdmac_chan	*atchan = (struct at_xdmac_chan *)data;
+	struct at_xdmac_desc	*desc;
+	u32			error_mask;
+
+	dev_dbg(chan2dev(&atchan->chan), "%s: status=0x%08lx\n",
+		 __func__, atchan->status);
+
+	error_mask = AT_XDMAC_CIS_RBEIS
+		     | AT_XDMAC_CIS_WBEIS
+		     | AT_XDMAC_CIS_ROIS;
+
+	if (at_xdmac_chan_is_cyclic(atchan)) {
+		at_xdmac_handle_cyclic(atchan);
+	} else if ((atchan->status & AT_XDMAC_CIS_LIS)
+		   || (atchan->status & error_mask)) {
+		struct dma_async_tx_descriptor  *txd;
+
+		if (atchan->status & AT_XDMAC_CIS_RBEIS)
+			dev_err(chan2dev(&atchan->chan), "read bus error!!!");
+		else if (atchan->status & AT_XDMAC_CIS_WBEIS)
+			dev_err(chan2dev(&atchan->chan), "write bus error!!!");
+		else if (atchan->status & AT_XDMAC_CIS_ROIS)
+			dev_err(chan2dev(&atchan->chan), "request overflow error!!!");
+
+		desc = list_first_entry(&atchan->xfers_list,
+					struct at_xdmac_desc,
+					xfer_node);
+		dev_vdbg(chan2dev(&atchan->chan), "%s: desc 0x%p\n", __func__, desc);
+		BUG_ON(!desc->active_xfer);
+
+		txd = &desc->tx_dma_desc;
+
+		at_xdmac_terminate_xfer(atchan, desc);
+
+		if (!at_xdmac_chan_is_cyclic(atchan)) {
+			dma_cookie_complete(txd);
+			if (txd->callback && (txd->flags & DMA_PREP_INTERRUPT))
+				txd->callback(txd->callback_param);
+		}
+
+		dma_run_dependencies(txd);
+
+		at_xdmac_advance_work(atchan);
+	}
+}
+
+static irqreturn_t at_xdmac_interrupt(int irq, void *dev_id)
+{
+	struct at_xdmac		*atxdmac = (struct at_xdmac *)dev_id;
+	struct at_xdmac_chan	*atchan;
+	u32 			imr, status, pending;
+	u32			chan_imr, chan_status;
+	int			i, ret = IRQ_NONE;
+
+	do {
+		imr = at_xdmac_read(atxdmac, AT_XDMAC_GIM);
+		status = at_xdmac_read(atxdmac, AT_XDMAC_GIS);
+		pending = status & imr;
+
+		dev_vdbg(atxdmac->dma.dev,
+			 "%s: status=0x%08x, imr=0x%08x, pending=0x%08x\n",
+			 __func__, status, imr, pending);
+
+		if (!pending)
+			break;
+
+		/* We have to find which channel has generated the interrupt. */
+		for (i = 0; i < atxdmac->dma.chancnt; i++) {
+			if (!((1 << i) & pending))
+				continue;
+
+			atchan = &atxdmac->chan[i];
+			chan_imr = at_xdmac_chan_read(atchan, AT_XDMAC_CIM);
+			chan_status = at_xdmac_chan_read(atchan, AT_XDMAC_CIS);
+			atchan->status = chan_status & chan_imr;
+			dev_vdbg(atxdmac->dma.dev,
+				 "%s: chan%d: imr=0x%x, status=0x%x\n",
+				 __func__, i, chan_imr, chan_status);
+			dev_vdbg(chan2dev(&atchan->chan),
+				 "%s: XDMAC_CC=0x%08x XDMAC_CNDA=0x%08x, "
+				 "XDMAC_CNDC=0x%08x, XDMAC_CSA=0x%08x, "
+				 "XDMAC_CDA=0x%08x, XDMAC_CUBC=0x%08x\n",
+				 __func__,
+				 at_xdmac_chan_read(atchan, AT_XDMAC_CC),
+				 at_xdmac_chan_read(atchan, AT_XDMAC_CNDA),
+				 at_xdmac_chan_read(atchan, AT_XDMAC_CNDC),
+				 at_xdmac_chan_read(atchan, AT_XDMAC_CSA),
+				 at_xdmac_chan_read(atchan, AT_XDMAC_CDA),
+				 at_xdmac_chan_read(atchan, AT_XDMAC_CUBC));
+
+			if (atchan->status & (AT_XDMAC_CIS_RBEIS | AT_XDMAC_CIS_WBEIS))
+				at_xdmac_write(atxdmac, AT_XDMAC_GD, atchan->mask);
+
+			tasklet_schedule(&atchan->tasklet);
+			ret = IRQ_HANDLED;
+		}
+
+	} while (pending);
+
+	return ret;
+}
+
+static void at_xdmac_issue_pending(struct dma_chan *chan)
+{
+	struct at_xdmac_chan *atchan = to_at_xdmac_chan(chan);
+
+	dev_dbg(chan2dev(&atchan->chan), "%s\n", __func__);
+
+	if (!at_xdmac_chan_is_cyclic(atchan))
+		at_xdmac_advance_work(atchan);
+
+	return;
+}
+
+static int at_xdmac_control(struct dma_chan *chan, enum dma_ctrl_cmd cmd,
+			    unsigned long arg)
+{
+	struct at_xdmac_desc	*desc, *_desc;
+	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(chan);
+	struct at_xdmac		*atxdmac = to_at_xdmac(atchan->chan.device);
+	unsigned long		flags;
+	int			ret = 0;
+
+	dev_dbg(chan2dev(chan), "%s: cmd=%d\n", __func__, cmd);
+
+	spin_lock_irqsave(&atchan->lock, flags);
+
+	switch (cmd) {
+	case DMA_PAUSE:
+		at_xdmac_write(atxdmac, AT_XDMAC_GRWS, atchan->mask);
+		set_bit(AT_XDMAC_CHAN_IS_PAUSED, &atchan->status);
+		break;
+	case DMA_RESUME:
+		if (!at_xdmac_chan_is_paused(atchan))
+			break;
+
+		at_xdmac_write(atxdmac, AT_XDMAC_GRWR, atchan->mask);
+		clear_bit(AT_XDMAC_CHAN_IS_PAUSED, &atchan->status);
+		break;
+	case DMA_TERMINATE_ALL:
+		at_xdmac_write(atxdmac, AT_XDMAC_GIE, atchan->mask);
+		at_xdmac_write(atxdmac, AT_XDMAC_GD, atchan->mask);
+		while (at_xdmac_read(atxdmac, AT_XDMAC_GS) & atchan->mask)
+			cpu_relax();
+
+		/* Cancel all pending transfers. */
+		list_for_each_entry_safe(desc, _desc, &atchan->xfers_list, xfer_node)
+			at_xdmac_terminate_xfer(atchan, desc);
+
+		clear_bit(AT_XDMAC_CHAN_IS_CYCLIC, &atchan->status);
+		break;
+	case DMA_SLAVE_CONFIG:
+		ret = at_xdmac_set_slave_config(chan,
+				(struct dma_slave_config *)arg);
+		break;
+	default:
+		dev_err(chan2dev(chan),
+			"unmanaged or unknown dma control cmd: %d\n", cmd);
+		ret = -ENXIO;
+	}
+
+	spin_unlock_irqrestore(&atchan->lock, flags);
+
+	return ret;
+}
+
+static int at_xdmac_alloc_chan_resources(struct dma_chan *chan)
+{
+	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(chan);
+	struct at_xdmac_desc	*desc;
+	unsigned long		flags;
+	int			i;
+
+	spin_lock_irqsave(&atchan->lock, flags);
+
+	if (at_xdmac_chan_is_enabled(atchan)) {
+		dev_err(chan2dev(chan),
+			"can't allocate channel resources (channel enabled)\n");
+		i = -EIO;
+		goto spin_unlock;
+	}
+
+	if (!list_empty(&atchan->free_descs_list)) {
+		dev_err(chan2dev(chan),
+			"can't allocate channel resources (channel not free from a previous use)\n");
+		i = -EIO;
+		goto spin_unlock;
+	}
+
+	for (i = 0; i < init_nr_desc_per_channel; i++) {
+		desc = at_xdmac_alloc_desc(chan, GFP_ATOMIC);
+		if (!desc) {
+			dev_warn(chan2dev(chan),
+				"only %d descriptors have been allocated\n", i);
+			break;
+		}
+		list_add_tail(&desc->desc_node, &atchan->free_descs_list);
+	}
+
+	dma_cookie_init(chan);
+
+	dev_dbg(chan2dev(chan), "%s: allocated %d descriptors\n", __func__, i);
+
+spin_unlock:
+	spin_unlock_irqrestore(&atchan->lock, flags);
+	return i;
+}
+
+static void at_xdmac_free_chan_resources(struct dma_chan *chan)
+{
+	struct at_xdmac_chan	*atchan = to_at_xdmac_chan(chan);
+	struct at_xdmac		*atxdmac = to_at_xdmac(chan->device);
+	struct at_xdmac_desc	*desc, *_desc;
+
+	list_for_each_entry_safe(desc, _desc, &atchan->free_descs_list, desc_node) {
+		dev_dbg(chan2dev(chan), "%s: freeing descriptor %p\n", __func__, desc);
+		list_del(&desc->desc_node);
+		dma_pool_free(atxdmac->at_xdmac_desc_pool, desc, desc->tx_dma_desc.phys);
+	}
+
+	return;
+}
+
+static int atmel_xdmac_prepare(struct device *dev)
+{
+	struct platform_device	*pdev = to_platform_device(dev);
+	struct at_xdmac		*atxdmac = platform_get_drvdata(pdev);
+	struct dma_chan		*chan, *_chan;
+
+	list_for_each_entry_safe(chan, _chan, &atxdmac->dma.channels, device_node) {
+		struct at_xdmac_chan	*atchan = to_at_xdmac_chan(chan);
+
+		/* Wait for transfer completion, except in cyclic case. */
+		if (at_xdmac_chan_is_enabled(atchan) && !at_xdmac_chan_is_cyclic(atchan))
+			return -EAGAIN;
+	}
+	return 0;
+}
+
+static int atmel_xdmac_suspend_noirq(struct device *dev)
+{
+	struct platform_device	*pdev = to_platform_device(dev);
+	struct at_xdmac		*atxdmac = platform_get_drvdata(pdev);
+	struct dma_chan		*chan, *_chan;
+
+	list_for_each_entry_safe(chan, _chan, &atxdmac->dma.channels, device_node) {
+		struct at_xdmac_chan	*atchan = to_at_xdmac_chan(chan);
+
+		if (at_xdmac_chan_is_cyclic(atchan)) {
+			if (!at_xdmac_chan_is_paused(atchan))
+				at_xdmac_control(chan, DMA_PAUSE, 0);
+			atchan->save_cim = at_xdmac_chan_read(atchan, AT_XDMAC_CIM);
+			atchan->save_cnda = at_xdmac_chan_read(atchan, AT_XDMAC_CNDA);
+			atchan->save_cndc = at_xdmac_chan_read(atchan, AT_XDMAC_CNDC);
+		}
+	}
+	atxdmac->save_gim = at_xdmac_read(atxdmac, AT_XDMAC_GIM);
+
+	at_xdmac_off(atxdmac);
+	clk_disable_unprepare(atxdmac->clk);
+	return 0;
+}
+
+static int atmel_xdmac_resume_noirq(struct device *dev)
+{
+	struct platform_device	*pdev = to_platform_device(dev);
+	struct at_xdmac		*atxdmac = platform_get_drvdata(pdev);
+	struct at_xdmac_chan	*atchan;
+	struct dma_chan		*chan, *_chan;
+	int			i;
+
+	clk_prepare_enable(atxdmac->clk);
+
+	/* Clear pending interrupts. */
+	for (i = 0; i < atxdmac->dma.chancnt; i++) {
+		atchan = &atxdmac->chan[i];
+		while (at_xdmac_chan_read(atchan, AT_XDMAC_CIS))
+			cpu_relax();
+	}
+
+	at_xdmac_write(atxdmac, AT_XDMAC_GIE, atxdmac->save_gim);
+	at_xdmac_write(atxdmac, AT_XDMAC_GE, atxdmac->save_gs);
+	list_for_each_entry_safe(chan, _chan, &atxdmac->dma.channels, device_node) {
+		atchan = to_at_xdmac_chan(chan);
+		at_xdmac_chan_write(atchan, AT_XDMAC_CC, atchan->cfg);
+		if (at_xdmac_chan_is_cyclic(atchan)) {
+			at_xdmac_chan_write(atchan, AT_XDMAC_CNDA, atchan->save_cnda);
+			at_xdmac_chan_write(atchan, AT_XDMAC_CNDC, atchan->save_cndc);
+			at_xdmac_chan_write(atchan, AT_XDMAC_CIE, atchan->save_cim);
+			at_xdmac_write(atxdmac, AT_XDMAC_GE, atchan->mask);
+		}
+	}
+	return 0;
+}
+
+static int at_xdmac_probe(struct platform_device *pdev)
+{
+	struct resource	*res;
+	struct at_xdmac	*atxdmac;
+	int		irq, size, nr_channels, i, ret;
+	void __iomem	*base;
+	u32		reg;
+
+	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
+	if (!res)
+		return -EINVAL;
+
+	irq = platform_get_irq(pdev, 0);
+	if (irq < 0)
+		return irq;
+
+	base = devm_ioremap_resource(&pdev->dev, res);
+	if (IS_ERR(base))
+		return PTR_ERR(base);
+
+	/*
+	 * Read number of xdmac channels, read helper function can't be used
+	 * since atxdmac is not yet allocated and we need to know the number
+	 * of channels to do the allocation.
+	 */
+	reg = readl_relaxed(base + AT_XDMAC_GTYPE);
+	nr_channels = AT_XDMAC_NB_CH(reg);
+	if (nr_channels > AT_XDMAC_MAX_CHAN) {
+		dev_err(&pdev->dev, "invalid number of channels (%u)\n",
+			nr_channels);
+		return -EINVAL;
+	}
+
+	size = sizeof(*atxdmac);
+	size += nr_channels * sizeof(struct at_xdmac_chan);
+	atxdmac = devm_kzalloc(&pdev->dev, size, GFP_KERNEL);
+	if (!atxdmac) {
+		dev_err(&pdev->dev, "can't allocate at_xdmac structure\n");
+		return -ENOMEM;
+	}
+
+	atxdmac->regs = base;
+	atxdmac->irq = irq;
+
+	ret = devm_request_irq(&pdev->dev, atxdmac->irq, at_xdmac_interrupt, 0,
+			       "at_xdmac", atxdmac);
+	if (ret) {
+		dev_err(&pdev->dev, "can't request irq\n");
+		return ret;
+	}
+
+	atxdmac->clk = devm_clk_get(&pdev->dev, "dma_clk");
+	if (IS_ERR(atxdmac->clk)) {
+		dev_err(&pdev->dev, "can't get dma_clk\n");
+		return PTR_ERR(atxdmac->clk);
+	}
+
+	ret = clk_prepare_enable(atxdmac->clk);
+	if (ret) {
+		dev_err(&pdev->dev, "can't prepare or enable clock\n");
+		return ret;
+	}
+
+	atxdmac->at_xdmac_desc_pool =
+		dmam_pool_create(dev_name(&pdev->dev), &pdev->dev,
+				sizeof(struct at_xdmac_desc), 4, 0);
+	if (!atxdmac->at_xdmac_desc_pool) {
+		dev_err(&pdev->dev, "no memory for descriptors dma pool\n");
+		ret = -ENOMEM;
+		goto err_clk_disable;
+	}
+
+	dma_cap_set(DMA_CYCLIC, atxdmac->dma.cap_mask);
+	dma_cap_set(DMA_MEMCPY, atxdmac->dma.cap_mask);
+	dma_cap_set(DMA_SLAVE, atxdmac->dma.cap_mask);
+	atxdmac->dma.dev				= &pdev->dev;
+	atxdmac->dma.device_alloc_chan_resources	= at_xdmac_alloc_chan_resources;
+	atxdmac->dma.device_free_chan_resources		= at_xdmac_free_chan_resources;
+	atxdmac->dma.device_tx_status			= at_xdmac_tx_status;
+	atxdmac->dma.device_issue_pending		= at_xdmac_issue_pending;
+	atxdmac->dma.device_prep_dma_cyclic		= at_xdmac_prep_dma_cyclic;
+	atxdmac->dma.device_prep_dma_memcpy		= at_xdmac_prep_dma_memcpy;
+	atxdmac->dma.device_prep_slave_sg		= at_xdmac_prep_slave_sg;
+	atxdmac->dma.device_control			= at_xdmac_control;
+	atxdmac->dma.chancnt				= nr_channels;
+
+	/* Disable all chans and interrupts. */
+	at_xdmac_off(atxdmac);
+
+	/* Init channels. */
+	INIT_LIST_HEAD(&atxdmac->dma.channels);
+	for (i = 0; i < nr_channels; i++) {
+		struct at_xdmac_chan *atchan = &atxdmac->chan[i];
+
+		atchan->chan.device = &atxdmac->dma;
+		list_add_tail(&atchan->chan.device_node,
+			      &atxdmac->dma.channels);
+
+		atchan->ch_regs = at_xdmac_chan_reg_base(atxdmac, i);
+		atchan->mask = 1 << i;
+
+		spin_lock_init(&atchan->lock);
+		INIT_LIST_HEAD(&atchan->xfers_list);
+		INIT_LIST_HEAD(&atchan->free_descs_list);
+		tasklet_init(&atchan->tasklet, at_xdmac_tasklet,
+			     (unsigned long)atchan);
+
+		/* Clear pending interrupts. */
+		while (at_xdmac_chan_read(atchan, AT_XDMAC_CIS))
+			cpu_relax();
+	}
+	platform_set_drvdata(pdev, atxdmac);
+
+	ret = dma_async_device_register(&atxdmac->dma);
+	if (ret) {
+		dev_err(&pdev->dev, "fail to register DMA engine device\n");
+		goto err_clk_disable;
+	}
+
+	ret = of_dma_controller_register(pdev->dev.of_node,
+					 at_xdmac_xlate, atxdmac);
+	if (ret) {
+		dev_err(&pdev->dev, "could not register of dma controller\n");
+		goto err_dma_unregister;
+	}
+
+	dev_info(&pdev->dev, "%d channels, mapped at 0x%p\n",
+		 nr_channels, atxdmac->regs);
+
+	return 0;
+
+err_dma_unregister:
+	dma_async_device_unregister(&atxdmac->dma);
+err_clk_disable:
+	clk_disable_unprepare(atxdmac->clk);
+	return ret;
+}
+
+static int at_xdmac_remove(struct platform_device *pdev)
+{
+	struct at_xdmac	*atxdmac = (struct at_xdmac*)platform_get_drvdata(pdev);
+	int		i;
+
+	at_xdmac_off(atxdmac);
+	of_dma_controller_free(pdev->dev.of_node);
+	dma_async_device_unregister(&atxdmac->dma);
+	clk_disable_unprepare(atxdmac->clk);
+
+	synchronize_irq(atxdmac->irq);
+
+	for (i = 0; i < atxdmac->dma.chancnt; i++) {
+		struct at_xdmac_chan *atchan = &atxdmac->chan[i];
+
+		tasklet_kill(&atchan->tasklet);
+		at_xdmac_free_chan_resources(&atchan->chan);
+	}
+
+	return 0;
+}
+
+static const struct dev_pm_ops atmel_xdmac_dev_pm_ops = {
+	.prepare	= atmel_xdmac_prepare,
+	.suspend_noirq	= atmel_xdmac_suspend_noirq,
+	.resume_noirq	= atmel_xdmac_resume_noirq,
+};
+
+static const struct of_device_id atmel_xdmac_dt_ids[] = {
+	{
+		.compatible = "atmel,sama5d4-dma",
+	}, {
+		/* sentinel */
+	}
+};
+MODULE_DEVICE_TABLE(of, atmel_xdmac_dt_ids);
+
+static struct platform_driver at_xdmac_driver = {
+	.probe		= at_xdmac_probe,
+	.remove		= at_xdmac_remove,
+	.driver = {
+		.name		= "at_xdmac",
+		.owner		= THIS_MODULE,
+		.of_match_table	= of_match_ptr(atmel_xdmac_dt_ids),
+		.pm		= &atmel_xdmac_dev_pm_ops,
+	}
+};
+
+static int __init at_xdmac_init(void)
+{
+	return platform_driver_probe(&at_xdmac_driver, at_xdmac_probe);
+}
+subsys_initcall(at_xdmac_init);
+
+MODULE_DESCRIPTION("Atmel Extended DMA Controller driver");
+MODULE_AUTHOR("Ludovic Desroches <ludovic.desroches@atmel.com>");
+MODULE_LICENSE("GPL");
diff --git a/drivers/dma/at_xdmac.h b/drivers/dma/at_xdmac.h
new file mode 100644
index 0000000..b33c008
--- /dev/null
+++ b/drivers/dma/at_xdmac.h
@@ -0,0 +1,308 @@ 
+/*
+ * Copyright (C) 2014 Atmel Corporation
+ *
+ * Author: Ludovic Desroches <ludovic.desroches@atmel.com>
+ *
+ * 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.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef __AT_XDMAC_H__
+#define __AT_XDMAC_H__
+
+#include <linux/clk.h>
+#include <linux/dmaengine.h>
+#include <linux/dmapool.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/of_dma.h>
+#include <linux/platform_device.h>
+
+#include "dmaengine.h"
+
+/* Global registers */
+#define AT_XDMAC_GTYPE		0x00	/* Global Type Register */
+#define		AT_XDMAC_NB_CH(i)	(((i) & 0x1F) + 1)		/* Number of Channels Minus One */
+#define		AT_XDMAC_FIFO_SZ(i)	(((i) >> 5) & 0x7FF)		/* Number of Bytes */
+#define		AT_XDMAC_NB_REQ(i)	((((i) >> 16) & 0x3F) + 1)	/* Number of Peripheral Requests Minus One */
+#define AT_XDMAC_GCFG		0x04	/* Global Configuration Register */
+#define AT_XDMAC_GWAC		0x08	/* Global Weighted Arbiter Configuration Register */
+#define AT_XDMAC_GIE		0x0C	/* Global Interrupt Enable Register */
+#define AT_XDMAC_GID		0x10	/* Global Interrupt Disable Register */
+#define AT_XDMAC_GIM		0x14	/* Global Interrupt Mask Register */
+#define AT_XDMAC_GIS		0x18	/* Global Interrupt Status Register */
+#define AT_XDMAC_GE		0x1C	/* Global Channel Enable Register */
+#define AT_XDMAC_GD		0x20	/* Global Channel Disable Register */
+#define AT_XDMAC_GS		0x24	/* Global Channel Status Register */
+#define AT_XDMAC_GRS		0x28	/* Global Channel Read Suspend Register */
+#define AT_XDMAC_GWS		0x2C	/* Global Write Suspend Register */
+#define AT_XDMAC_GRWS		0x30	/* Global Channel Read Write Suspend Register */
+#define AT_XDMAC_GRWR		0x34	/* Global Channel Read Write Resume Register */
+#define AT_XDMAC_GSWR		0x38	/* Global Channel Software Request Register */
+#define AT_XDMAC_GSWS		0x3C	/* Global channel Software Request Status Register */
+#define AT_XDMAC_GSWF		0x40	/* Global Channel Software Flush Request Register */
+#define AT_XDMAC_VERSION	0xFFC	/* XDMAC Version Register */
+
+/* Channel relative registers offsets */
+#define AT_XDMAC_CIE		0x00	/* Channel Interrupt Enable Register */
+#define		AT_XDMAC_CIE_BIE	BIT(0)	/* End of Block Interrupt Enable Bit */
+#define		AT_XDMAC_CIE_LIE	BIT(1)	/* End of Linked List Interrupt Enable Bit */
+#define		AT_XDMAC_CIE_DIE	BIT(2)	/* End of Disable Interrupt Enable Bit */
+#define		AT_XDMAC_CIE_FIE	BIT(3)	/* End of Flush Interrupt Enable Bit */
+#define		AT_XDMAC_CIE_RBEIE	BIT(4)	/* Read Bus Error Interrupt Enable Bit */
+#define		AT_XDMAC_CIE_WBEIE	BIT(5)	/* Write Bus Error Interrupt Enable Bit */
+#define		AT_XDMAC_CIE_ROIE	BIT(6)	/* Request Overflow Interrupt Enable Bit */
+#define AT_XDMAC_CID		0x04	/* Channel Interrupt Disable Register */
+#define		AT_XDMAC_CID_BID	BIT(0)	/* End of Block Interrupt Disable Bit */
+#define		AT_XDMAC_CID_LID	BIT(1)	/* End of Linked List Interrupt Disable Bit */
+#define		AT_XDMAC_CID_DID	BIT(2)	/* End of Disable Interrupt Disable Bit */
+#define		AT_XDMAC_CID_FID	BIT(3)	/* End of Flush Interrupt Disable Bit */
+#define		AT_XDMAC_CID_RBEID	BIT(4)	/* Read Bus Error Interrupt Disable Bit */
+#define		AT_XDMAC_CID_WBEID	BIT(5)	/* Write Bus Error Interrupt Disable Bit */
+#define		AT_XDMAC_CID_ROID	BIT(6)	/* Request Overflow Interrupt Disable Bit */
+#define AT_XDMAC_CIM		0x08	/* Channel Interrupt Mask Register */
+#define		AT_XDMAC_CIM_BIM	BIT(0)	/* End of Block Interrupt Mask Bit */
+#define		AT_XDMAC_CIM_LIM	BIT(1)	/* End of Linked List Interrupt Mask Bit */
+#define		AT_XDMAC_CIM_DIM	BIT(2)	/* End of Disable Interrupt Mask Bit */
+#define		AT_XDMAC_CIM_FIM	BIT(3)	/* End of Flush Interrupt Mask Bit */
+#define		AT_XDMAC_CIM_RBEIM	BIT(4)	/* Read Bus Error Interrupt Mask Bit */
+#define		AT_XDMAC_CIM_WBEIM	BIT(5)	/* Write Bus Error Interrupt Mask Bit */
+#define		AT_XDMAC_CIM_ROIM	BIT(6)	/* Request Overflow Interrupt Mask Bit */
+#define AT_XDMAC_CIS		0x0C	/* Channel Interrupt Status Register */
+#define		AT_XDMAC_CIS_BIS	BIT(0)	/* End of Block Interrupt Status Bit */
+#define		AT_XDMAC_CIS_LIS	BIT(1)	/* End of Linked List Interrupt Status Bit */
+#define		AT_XDMAC_CIS_DIS	BIT(2)	/* End of Disable Interrupt Status Bit */
+#define		AT_XDMAC_CIS_FIS	BIT(3)	/* End of Flush Interrupt Status Bit */
+#define		AT_XDMAC_CIS_RBEIS	BIT(4)	/* Read Bus Error Interrupt Status Bit */
+#define		AT_XDMAC_CIS_WBEIS	BIT(5)	/* Write Bus Error Interrupt Status Bit */
+#define		AT_XDMAC_CIS_ROIS	BIT(6)	/* Request Overflow Interrupt Status Bit */
+#define AT_XDMAC_CSA		0x10	/* Channel Source Address Register */
+#define AT_XDMAC_CDA		0x14	/* Channel Destination Address Register */
+#define AT_XDMAC_CNDA		0x18	/* Channel Next Descriptor Address Register */
+#define 	AT_XDMAC_CNDA_NDAIF(i)	((i) & 0x1)			/* Channel x Next Descriptor Interface */
+#define		AT_XDMAC_CNDA_NDA(i)	((i) & 0xfffffffc)		/* Channel x Next Descriptor Address */
+#define AT_XDMAC_CNDC		0x1C	/* Channel Next Descriptor Control Register */
+#define		AT_XDMAC_CNDC_NDE		(0x1 << 0)		/* Channel x Next Descriptor Enable */
+#define		AT_XDMAC_CNDC_NDSUP		(0x1 << 1)		/* Channel x Next Descriptor Source Update */
+#define		AT_XDMAC_CNDC_NDDUP		(0x1 << 2)		/* Channel x Next Descriptor Destination Update */
+#define		AT_XDMAC_CNDC_NDVIEW_NDV0	(0x0 << 3)		/* Channel x Next Descriptor View 0 */
+#define		AT_XDMAC_CNDC_NDVIEW_NDV1	(0x1 << 3)		/* Channel x Next Descriptor View 1 */
+#define		AT_XDMAC_CNDC_NDVIEW_NDV2	(0x2 << 3)		/* Channel x Next Descriptor View 2 */
+#define		AT_XDMAC_CNDC_NDVIEW_NDV3	(0x3 << 3)		/* Channel x Next Descriptor View 3 */
+#define AT_XDMAC_CUBC		0x20	/* Channel Microblock Control Register */
+#define AT_XDMAC_CBC		0x24	/* Channel Block Control Register */
+#define AT_XDMAC_CC		0x28	/* Channel Configuration Register */
+#define		AT_XDMAC_CC_TYPE	(0x1 << 0)	/* Channel Transfer Type */
+#define			AT_XDMAC_CC_TYPE_MEM_TRAN	(0x0 << 0)	/* Memory to Memory Transfer */
+#define			AT_XDMAC_CC_TYPE_PER_TRAN	(0x1 << 0)	/* Peripheral to Memory or Memory to Peripheral Transfer */
+#define		AT_XDMAC_CC_MBSIZE_MASK	(0x3 << 1)
+#define			AT_XDMAC_CC_MBSIZE_SINGLE	(0x0 << 1)
+#define			AT_XDMAC_CC_MBSIZE_FOUR		(0x1 << 1)
+#define			AT_XDMAC_CC_MBSIZE_EIGHT	(0x2 << 1)
+#define			AT_XDMAC_CC_MBSIZE_SIXTEEN	(0x3 << 1)
+#define		AT_XDMAC_CC_DSYNC	(0x1 << 4)	/* Channel Synchronization */
+#define			AT_XDMAC_CC_DSYNC_PER2MEM	(0x0 << 4)
+#define			AT_XDMAC_CC_DSYNC_MEM2PER	(0x1 << 4)
+#define		AT_XDMAC_CC_PROT	(0x1 << 5)	/* Channel Protection */
+#define			AT_XDMAC_CC_PROT_SEC		(0x0 << 5)
+#define			AT_XDMAC_CC_PROT_UNSEC		(0x1 << 5)
+#define		AT_XDMAC_CC_SWREQ	(0x1 << 6)	/* Channel Software Request Trigger */
+#define			AT_XDMAC_CC_SWREQ_HWR_CONNECTED	(0x0 << 6)
+#define			AT_XDMAC_CC_SWREQ_SWR_CONNECTED	(0x1 << 6)
+#define		AT_XDMAC_CC_MEMSET	(0x1 << 7)	/* Channel Fill Block of memory */
+#define			AT_XDMAC_CC_MEMSET_NORMAL_MODE	(0x0 << 7)
+#define			AT_XDMAC_CC_MEMSET_HW_MODE	(0x1 << 7)
+#define		AT_XDMAC_CC_CSIZE_MASK	(0x7 << 8)	/* Channel Chunk Size */
+#define			AT_XDMAC_CC_CSIZE_CHK_1		(0x0 << 8)
+#define			AT_XDMAC_CC_CSIZE_CHK_2		(0x1 << 8)
+#define			AT_XDMAC_CC_CSIZE_CHK_4		(0x2 << 8)
+#define			AT_XDMAC_CC_CSIZE_CHK_8		(0x3 << 8)
+#define			AT_XDMAC_CC_CSIZE_CHK_16	(0x4 << 8)
+#define		AT_XDMAC_CC_DWIDTH(i)	((i) << 11)	/* Channel Data Width */
+#define			AT_XDMAC_CC_DWIDTH_BYTE		0x0
+#define			AT_XDMAC_CC_DWIDTH_HALFWORD	0x1
+#define			AT_XDMAC_CC_DWIDTH_WORD		0x2
+#define			AT_XDMAC_CC_DWIDTH_DWORD	0x3
+#define		AT_XDMAC_CC_SIF(i)	((0x1 & (i)) << 13)	/* Channel Source Interface Identifier */
+#define		AT_XDMAC_CC_DIF(i)	((0x1 & (i)) << 14)	/* Channel Destination Interface Identifier */
+#define		AT_XDMAC_CC_SAM_MASK	(0x3 << 16)	/* Channel Source Addressing Mode */
+#define			AT_XDMAC_CC_SAM_FIXED_AM	(0x0 << 16)
+#define			AT_XDMAC_CC_SAM_INCREMENTED_AM	(0x1 << 16)
+#define			AT_XDMAC_CC_SAM_UBS_AM		(0x2 << 16)
+#define			AT_XDMAC_CC_SAM_UBS_DS_AM	(0x3 << 16)
+#define		AT_XDMAC_CC_DAM_MASK	(0x3 << 18)	/* Channel Source Addressing Mode */
+#define			AT_XDMAC_CC_DAM_FIXED_AM	(0x0 << 18)
+#define			AT_XDMAC_CC_DAM_INCREMENTED_AM	(0x1 << 18)
+#define			AT_XDMAC_CC_DAM_UBS_AM		(0x2 << 18)
+#define			AT_XDMAC_CC_DAM_UBS_DS_AM	(0x3 << 18)
+#define		AT_XDMAC_CC_INITD	(0x1 << 21)	/* Channel Initialization Terminated (read only) */
+#define			AT_XDMAC_CC_INITD_TERMINATED	(0x0 << 21)
+#define			AT_XDMAC_CC_INITD_IN_PROGRESS	(0x1 << 21)
+#define		AT_XDMAC_CC_RDIP	(0x1 << 22)	/* Read in Progress (read only) */
+#define			AT_XDMAC_CC_RDIP_DONE		(0x0 << 22)
+#define			AT_XDMAC_CC_RDIP_IN_PROGRESS	(0x1 << 22)
+#define		AT_XDMAC_CC_WDIP	(0x1 << 23)	/* Write in Progress (read only) */
+#define			AT_XDMAC_CC_WDIP_DONE		(0x0 << 23)
+#define			AT_XDMAC_CC_WDIP_IN_PROGRESS	(0x1 << 23)
+#define		AT_XDMAC_CC_PERID(i)	(0x7f & (h) << 24)	/* Channel Peripheral Identifier */
+#define AT_XDMAC_CDS_MSP	0x2C	/* Channel Data Stride Memory Set Pattern */
+#define AT_XDMAC_CSUS		0x30	/* Channel Source Microblock Stride */
+#define AT_XDMAC_CDUS		0x34	/* Channel Destination Microblock Stride */
+
+#define AT_XDMAC_CHAN_REG_BASE	0x50	/* Channel registers base address */
+
+/* Microblock control members */
+#define AT_XDMAC_MBR_UBC_UBLEN_MAX	0xFFFFFFUL	/* Maximum Microblock Length */
+#define AT_XDMAC_MBR_UBC_NDE		(0x1 << 24)	/* Next Descriptor Enable */
+#define AT_XDMAC_MBR_UBC_NSEN		(0x1 << 25)	/* Next Descriptor Source Update */
+#define AT_XDMAC_MBR_UBC_NDEN		(0x1 << 26)	/* Next Descriptor Destination Update */
+#define AT_XDMAC_MBR_UBC_NDV0		(0x0 << 27)	/* Next Descriptor View 0 */
+#define AT_XDMAC_MBR_UBC_NDV1		(0x1 << 27)	/* Next Descriptor View 1 */
+#define AT_XDMAC_MBR_UBC_NDV2		(0x2 << 27)	/* Next Descriptor View 2 */
+#define AT_XDMAC_MBR_UBC_NDV3		(0x3 << 27)	/* Next Descriptor View 3 */
+
+#define AT_XDMAC_MAX_CHAN	0x20
+
+enum atc_status {
+	AT_XDMAC_CHAN_IS_CYCLIC = 0,
+	AT_XDMAC_CHAN_IS_PAUSED,
+};
+
+/* ----- Channels ----- */
+struct at_xdmac_chan {
+	struct dma_chan		chan;
+	void __iomem		*ch_regs;
+	u32			mask;		/* Channel Mask */
+	u32			cfg;		/* Channel Configuration Register */
+	u8			perid;		/* Peripheral ID */
+	u8			dwidth;		/* Data Width */
+	u8			perif;		/* Peripheral Interface */
+	u8			memif;		/* Memory Interface */
+	u32			save_cim;
+	u32			save_cnda;
+	u32			save_cndc;
+	unsigned long		status;
+	struct tasklet_struct	tasklet;
+	struct dma_slave_config	dma_sconfig;
+
+	spinlock_t		lock;
+
+	struct list_head	xfers_list;
+	struct list_head	free_descs_list;
+};
+
+
+/* ----- Controller ----- */
+struct at_xdmac {
+	struct dma_device	dma;
+	void __iomem		*regs;
+	int			irq;
+	struct clk		*clk;
+	u32			save_gim;
+	u32			save_gs;
+	struct dma_pool		*at_xdmac_desc_pool;
+	struct at_xdmac_chan	chan[0];
+};
+
+
+/* ----- Descriptors ----- */
+
+/* Linked List Descriptor */
+struct at_xdmac_lld {
+	dma_addr_t	mbr_nda;	/* Next Descriptor Member */
+	u32		mbr_ubc;	/* Microblock Control Member */
+	dma_addr_t	mbr_sa;		/* Source Address Member */
+	dma_addr_t	mbr_da;		/* Destination Address Member */
+	u32		mbr_cfg;	/* Configuration Register */
+};
+
+
+struct at_xdmac_desc {
+	struct at_xdmac_lld		lld;
+	enum dma_transfer_direction	direction;
+	struct dma_async_tx_descriptor	tx_dma_desc;
+	struct list_head		desc_node;
+	/* Following members are only used by the first descriptor */
+	bool				active_xfer;
+	unsigned int			xfer_size;
+	struct list_head		descs_list;
+	struct list_head		xfer_node;
+};
+
+static inline void __iomem *at_xdmac_chan_reg_base(struct at_xdmac *atxdmac, unsigned int chan_nb)
+{
+	return (void __iomem*)(atxdmac->regs + (AT_XDMAC_CHAN_REG_BASE + chan_nb * 0x40));
+}
+
+#define at_xdmac_read(atxdmac, reg) readl_relaxed((atxdmac)->regs + (reg))
+#define at_xdmac_write(atxdmac, reg, value) \
+	writel_relaxed((value), (atxdmac)->regs + (reg))
+
+#define at_xdmac_chan_read(atchan, reg) readl_relaxed((atchan)->ch_regs + (reg))
+#define at_xdmac_chan_write(atchan, reg, value) writel_relaxed((value), (atchan)->ch_regs + (reg))
+
+static inline struct at_xdmac_chan *to_at_xdmac_chan(struct dma_chan *dchan)
+{
+	return container_of(dchan, struct at_xdmac_chan, chan);
+}
+
+static struct device *chan2dev(struct dma_chan *chan)
+{
+	return &chan->dev->device;
+}
+
+static inline struct at_xdmac *to_at_xdmac(struct dma_device *ddev)
+{
+	return container_of(ddev, struct at_xdmac, dma);
+}
+
+static inline struct at_xdmac_desc *txd_to_at_desc(struct dma_async_tx_descriptor *txd)
+{
+	return container_of(txd, struct at_xdmac_desc, tx_dma_desc);
+}
+
+static inline int at_xdmac_chan_is_cyclic(struct at_xdmac_chan *atchan)
+{
+	return test_bit(AT_XDMAC_CHAN_IS_CYCLIC, &atchan->status);
+}
+
+static inline int at_xdmac_chan_is_paused(struct at_xdmac_chan *atchan)
+{
+	return test_bit(AT_XDMAC_CHAN_IS_PAUSED, &atchan->status);
+}
+
+static inline u32 at_xdmac_csize(u32 maxburst)
+{
+	u32 csize;
+
+	switch (ffs(maxburst) -1) {
+	case 0:
+		csize = AT_XDMAC_CC_CSIZE_CHK_1;
+		break;
+	case 1:
+		csize = AT_XDMAC_CC_CSIZE_CHK_2;
+		break;
+	case 2:
+		csize = AT_XDMAC_CC_CSIZE_CHK_4;
+		break;
+	case 3:
+		csize = AT_XDMAC_CC_CSIZE_CHK_8;
+		break;
+	default:
+		csize = AT_XDMAC_CC_CSIZE_CHK_16;
+	}
+
+	return csize;
+};
+#endif /* __AT_XDMAC_H__ */
diff --git a/include/dt-bindings/dma/at91.h b/include/dt-bindings/dma/at91.h
index e835037..4875a50 100644
--- a/include/dt-bindings/dma/at91.h
+++ b/include/dt-bindings/dma/at91.h
@@ -9,6 +9,8 @@ 
 #ifndef __DT_BINDINGS_AT91_DMA_H__
 #define __DT_BINDINGS_AT91_DMA_H__
 
+/* ---------- HDMAC ---------- */
+
 /*
  * Source and/or destination peripheral ID
  */
@@ -24,4 +26,27 @@ 
 #define AT91_DMA_CFG_FIFOCFG_ALAP	(0x1 << AT91_DMA_CFG_FIFOCFG_OFFSET)	/* largest defined AHB burst */
 #define AT91_DMA_CFG_FIFOCFG_ASAP	(0x2 << AT91_DMA_CFG_FIFOCFG_OFFSET)	/* single AHB access */
 
+
+/* ---------- XDMAC ---------- */
+#define AT91_XDMAC_DT_MEM_IF_MASK	(0x1)
+#define AT91_XDMAC_DT_MEM_IF_OFFSET	(16)
+#define AT91_XDMAC_DT_MEM_IF(mem_if)	(((mem_if) & AT91_XDMAC_DT_MEM_IF_MASK) \
+					<< AT91_XDMAC_DT_MEM_IF_OFFSET)
+#define AT91_XDMAC_DT_GET_MEM_IF(cfg)	(((cfg) >> AT91_XDMAC_DT_MEM_IF_OFFSET) \
+					& AT91_XDMAC_DT_MEM_IF_MASK)
+
+#define AT91_XDMAC_DT_PER_IF_MASK	(0x1)
+#define AT91_XDMAC_DT_PER_IF_OFFSET	(0)
+#define AT91_XDMAC_DT_PER_IF(per_if)	(((per_if) & AT91_XDMAC_DT_PER_IF_MASK) \
+					<< AT91_XDMAC_DT_PER_IF_OFFSET)
+#define AT91_XDMAC_DT_GET_PER_IF(cfg)	(((cfg) >> AT91_XDMAC_DT_PER_IF_OFFSET) \
+					& AT91_XDMAC_DT_PER_IF_MASK)
+
+#define AT91_XDMAC_DT_PERID_MASK	(0x7f)
+#define AT91_XDMAC_DT_PERID_OFFSET	(24)
+#define AT91_XDMAC_DT_PERID(perid)	(((perid) & AT91_XDMAC_DT_PERID_MASK) \
+					<< AT91_XDMAC_DT_PERID_OFFSET)
+#define AT91_XDMAC_DT_GET_PERID(cfg)	(((cfg) >> AT91_XDMAC_DT_PERID_OFFSET) \
+					& AT91_XDMAC_DT_PERID_MASK)
+
 #endif /* __DT_BINDINGS_AT91_DMA_H__ */