[alsa-devel] [PATCHv2] dmaengine: Add support for BCM2835.

[Russell King - ARM Linux]

On Mon, Nov 11, 2013 at 05:33:41PM +0100, Florian Meier wrote:
> Add support for DMA controller of BCM2835 as used in the Raspberry Pi.
> Currently it only supports cyclic DMA.

It's better, but still modelled too much on omap-dma.c.

The changes I suggest below get rid of bcm2835_sg entirely, and also
remove a number of other OMAP specific bits from your driver.

> +struct bcm2835_dmadev {
> +	struct dma_device ddev;
> +	spinlock_t lock;
> +	struct tasklet_struct task;
> +	struct list_head pending;
> +	uint32_t chans_available;
> +	void __iomem *dma_base;
> +	int *dma_irq_numbers;
> +};
> +
> +struct bcm2835_dma_cb {
> +	unsigned long info;
> +	unsigned long src;
> +	unsigned long dst;
> +	unsigned long length;
> +	unsigned long stride;
> +	unsigned long next;
> +	unsigned long pad[2];

These are passed to the hardware, so you should not be relying on
"unsigned long" being the correct size.  It's best to use a type which
has a well defined size such as uint32_t.

However, there's also concerns about the endian-ness of the hardware
which should be addressed here too.  Does the DMA hardware follow the
CPU endian (native endian) or is it always little endian?  If it's
always little endian, you might consider using "le32" as a type here,
and then using cpu_to_le32() / le32_to_cpu() as appropraite in the
driver when writing / reading these fields.

> +};
> +
> +struct bcm2835_chan {
> +	struct virt_dma_chan vc;
> +	struct list_head node;
> +
> +	struct dma_slave_config	cfg;
> +	unsigned dma_sig;
> +	bool cyclic;
> +	unsigned dreq;
> +
> +	int dma_ch;
> +	struct bcm2835_desc *desc;
> +	unsigned sgidx;

sgidx  won't be required.

> +
> +	void __iomem *dma_chan_base;
> +	int dma_irq_number;
> +	struct irqaction dma_irq_handler;
> +};
> +
> +struct bcm2835_sg {
> +	dma_addr_t addr;
> +	uint32_t en;		/* number of elements (24-bit) */
> +	uint32_t fn;		/* number of frames (16-bit) */
> +};
> +
> +struct bcm2835_desc {
> +	struct virt_dma_desc vd;
> +	enum dma_transfer_direction dir;
> +	dma_addr_t dev_addr;
> +
> +	uint8_t es;
> +
> +	unsigned int control_block_size;
> +	struct bcm2835_dma_cb *control_block_base;
> +	dma_addr_t control_block_base_phys;
> +
> +	unsigned sglen;

Replace with:
	unsigned frames;

> +	struct bcm2835_sg sg[0];
> +};
> +
> +#define BCM2835_DMA_CS		0x00
> +#define BCM2835_DMA_ADDR	0x04
> +#define BCM2835_DMA_SOURCE_AD	0x0c
> +#define BCM2835_DMA_DEST_AD	0x10
> +#define BCM2835_DMA_NEXTCB	0x1C
> +
> +/* DMA CS Control and Status bits */
> +#define BCM2835_DMA_ACTIVE	(1 << 0)
> +#define BCM2835_DMA_INT	(1 << 2)
> +#define BCM2835_DMA_ISPAUSED	(1 << 4)  /* Pause requested or not active */
> +#define BCM2835_DMA_ISHELD	(1 << 5)  /* Is held by DREQ flow control */
> +#define BCM2835_DMA_ERR	(1 << 8)
> +#define BCM2835_DMA_ABORT	(1 << 30) /* stop current CB, go to next, WO */
> +#define BCM2835_DMA_RESET	(1 << 31) /* WO, self clearing */
> +
> +#define BCM2835_DMA_INT_EN	(1 << 0)
> +#define BCM2835_DMA_D_INC	(1 << 4)
> +#define BCM2835_DMA_D_DREQ	(1 << 6)
> +#define BCM2835_DMA_S_INC	(1 << 8)
> +#define BCM2835_DMA_S_DREQ	(1 << 6)
> +
> +#define BCM2835_DMA_PER_MAP(x)	((x) << 16)
> +
> +#define BCM2835_DMA_DATA_TYPE_S8 1
> +#define BCM2835_DMA_DATA_TYPE_S16 2
> +#define BCM2835_DMA_DATA_TYPE_S32 4
> +#define BCM2835_DMA_DATA_TYPE_S128 16
> +
> +/* valid only for channels 0 - 14, 15 has its own base address */
> +#define BCM2835_DMA_CHAN(n)	((n)<<8) /* base address */
> +#define BCM2835_DMA_CHANIO(dma_base, n) ((dma_base)+BCM2835_DMA_CHAN(n))

One space each side of the + and << please.

> +
> +static inline struct bcm2835_dmadev *to_bcm2835_dma_dev(struct dma_device *d)
> +{
> +	return container_of(d, struct bcm2835_dmadev, ddev);
> +}
> +
> +static inline struct bcm2835_chan *to_bcm2835_dma_chan(struct dma_chan *c)
> +{
> +	return container_of(c, struct bcm2835_chan, vc.chan);
> +}
> +
> +static inline struct bcm2835_desc *to_bcm2835_dma_desc(
> +		struct dma_async_tx_descriptor *t)
> +{
> +	return container_of(t, struct bcm2835_desc, vd.tx);
> +}
> +
> +static void bcm2835_dma_desc_free(struct virt_dma_desc *vd)
> +{
> +	struct bcm2835_desc *desc = container_of(vd, struct bcm2835_desc, vd);
> +	dma_free_coherent(desc->vd.tx.chan->device->dev,
> +			desc->control_block_size,
> +			desc->control_block_base,
> +			desc->control_block_base_phys);
> +	kfree(desc);
> +}
> +
> +static int bcm2835_dma_abort(void __iomem *dma_chan_base)
> +{
> +	unsigned long int cs;
> +	int rc = 0;
> +
> +	cs = readl(dma_chan_base + BCM2835_DMA_CS);
> +
> +	if (BCM2835_DMA_ACTIVE & cs) {
> +		long int timeout = 10000;
> +
> +		/* write 0 to the active bit - pause the DMA */
> +		writel(0, dma_chan_base + BCM2835_DMA_CS);
> +
> +		/* wait for any current AXI transfer to complete */
> +		while (0 != (cs & BCM2835_DMA_ISPAUSED) && --timeout >= 0)
> +			cs = readl(dma_chan_base + BCM2835_DMA_CS);
> +
> +		if (0 != (cs & BCM2835_DMA_ISPAUSED)) {

The two 0 != 's are implied: in C, zero is false, non-zero is true.

> +			/* we'll un-pause when we set of our next DMA */
> +			rc = -ETIMEDOUT;
> +
> +		} else if (BCM2835_DMA_ACTIVE & cs) {
> +			/* terminate the control block chain */
> +			writel(0, dma_chan_base + BCM2835_DMA_NEXTCB);
> +
> +			/* abort the whole DMA */
> +			writel(BCM2835_DMA_ABORT | BCM2835_DMA_ACTIVE,
> +			       dma_chan_base + BCM2835_DMA_CS);
> +		}
> +	}
> +
> +	return rc;
> +}
> +
> +static void bcm2835_dma_start_sg(struct bcm2835_chan *c, struct bcm2835_desc *d,
> +		unsigned idx)

You're having to wrap this anyway, but you're leaving the first line
just about acceptable.  Better would be:

static void bcm2835_dma_start_sg(struct bcm2835_chan *c,
	struct bcm2835_desc *d, unsigned idx)

IMHO.

> +{
> +	dsb();	/* ARM data synchronization (push) operation */
> +
> +	writel(d->control_block_base_phys, c->dma_chan_base+BCM2835_DMA_ADDR);
> +	writel(BCM2835_DMA_ACTIVE, c->dma_chan_base+BCM2835_DMA_CS);

Again, one space each side of the + please.

> +}
> +
> +static void bcm2835_dma_start_desc(struct bcm2835_chan *c)
> +{
> +	struct virt_dma_desc *vd = vchan_next_desc(&c->vc);
> +	struct bcm2835_desc *d;
> +
> +	if (!vd) {
> +		c->desc = NULL;
> +		return;
> +	}
> +
> +	list_del(&vd->node);
> +
> +	c->desc = d = to_bcm2835_dma_desc(&vd->tx);
> +	c->sgidx = 0;

c->sgidx won't be needed...

> +
> +	bcm2835_dma_start_sg(c, d, 0);

Move the contents of bcm2835_dma_start_sg() into here.

> +}
> +
> +static irqreturn_t bcm2835_dma_callback(int irq, void *data)
> +{
> +	struct bcm2835_chan *c = data;
> +	struct bcm2835_desc *d;
> +	unsigned long flags;
> +
> +	spin_lock_irqsave(&c->vc.lock, flags);
> +
> +	/* acknowledge interrupt */
> +	writel(BCM2835_DMA_INT, c->dma_chan_base + BCM2835_DMA_CS);
> +
> +	d = c->desc;
> +
> +	if (d) {
> +		if (!c->cyclic) {
> +			if (++c->sgidx < d->sglen) {
> +				bcm2835_dma_start_sg(c, d, c->sgidx);
> +			} else {

You don't need the sgidx stuff nor sglen here.

> +				bcm2835_dma_start_desc(c);
> +				vchan_cookie_complete(&d->vd);
> +			}
> +		} else {
> +			vchan_cyclic_callback(&d->vd);
> +		}
> +	}
> +
> +	/* keep the DMA engine running */
> +	dsb(); /* ARM synchronization barrier */
> +	writel(BCM2835_DMA_ACTIVE, c->dma_chan_base + BCM2835_DMA_CS);
> +
> +	spin_unlock_irqrestore(&c->vc.lock, flags);
> +
> +	return IRQ_HANDLED;
> +}
> +
> +/*
> + * This callback schedules all pending channels.  We could be more
> + * clever here by postponing allocation of the real DMA channels to
> + * this point, and freeing them when our virtual channel becomes idle.
> + *
> + * We would then need to deal with 'all channels in-use'
> + */
> +static void bcm2835_dma_sched(unsigned long data)
> +{
> +	struct bcm2835_dmadev *d = (struct bcm2835_dmadev *)data;
> +	LIST_HEAD(head);
> +
> +	spin_lock_irq(&d->lock);
> +	list_splice_tail_init(&d->pending, &head);
> +	spin_unlock_irq(&d->lock);
> +
> +	while (!list_empty(&head)) {
> +		struct bcm2835_chan *c = list_first_entry(&head,
> +			struct bcm2835_chan, node);
> +
> +		spin_lock_irq(&c->vc.lock);
> +		list_del_init(&c->node);
> +		bcm2835_dma_start_desc(c);
> +		spin_unlock_irq(&c->vc.lock);
> +	}
> +}
> +
> +static int bcm2835_dma_alloc_chan_resources(struct dma_chan *chan)
> +{
> +	struct bcm2835_chan *c = to_bcm2835_dma_chan(chan);
> +	int ret;
> +	struct bcm2835_dmadev *d = to_bcm2835_dma_dev(chan->device);
> +	uint32_t chans = d->chans_available;
> +	int chanID = 0;
> +
> +	dev_dbg(c->vc.chan.device->dev,
> +			"allocating channel for %u\n", c->dma_sig);
> +
> +	/* do not use the FIQ and BULK channels */
> +	chans &= ~0xD;
> +
> +	if (chans) {
> +		/* return the ordinal of the first channel in the bitmap */
> +		while (chans != 0 && (chans & 1) == 0) {
> +			chans >>= 1;
> +			chanID++;
> +		}
> +
> +		/* claim the channel */
> +		d->chans_available &= ~(1 << chanID);

This allocation really needs some kind of locking to avoid two concurrent
calls (or even preempted calls) from getting the same channel.

> +
> +		c->dma_chan_base = BCM2835_DMA_CHANIO(d->dma_base, chanID);
> +		c->dma_irq_number = d->dma_irq_numbers[chanID];
> +		c->dma_ch = chanID;
> +	} else {
> +		return -ENOMEM;
> +	}
> +
> +	c->dma_irq_handler.name = "DMA engine IRQ handler";
> +	c->dma_irq_handler.flags = 0;
> +	c->dma_irq_handler.handler = bcm2835_dma_callback;
> +
> +	ret = request_irq(c->dma_irq_number,
> +			bcm2835_dma_callback, 0, "DMA IRQ", c);
> +	if (ret < 0)
> +		return ret;
> +
> +	return 0;
> +}
> +
> +static void bcm2835_dma_free_chan_resources(struct dma_chan *chan)
> +{
> +	struct bcm2835_chan *c = to_bcm2835_dma_chan(chan);
> +	struct bcm2835_dmadev *d = to_bcm2835_dma_dev(chan->device);
> +
> +	vchan_free_chan_resources(&c->vc);
> +	d->chans_available |= (1 << c->dma_ch);

Also needs locking to prevent races with the above allocation code.
BTW, "1 << c->dma_ch" is acceptable - the additional parens don't gain
you much.

> +	free_irq(c->dma_irq_number, c);
> +
> +	dev_dbg(c->vc.chan.device->dev, "freeing channel for %u\n", c->dma_sig);
> +}
> +
> +static size_t bcm2835_dma_sg_size(struct bcm2835_sg *sg)
> +{
> +	return sg->en * sg->fn;
> +}

The above function won't be needed...

> +
> +static size_t bcm2835_dma_desc_size(struct bcm2835_desc *d)
> +{
> +	unsigned i;
> +	size_t size;
> +
> +	for (size = i = 0; i < d->sglen; i++)
> +		size += bcm2835_dma_sg_size(&d->sg[i]);
> +
> +	return size * d->es;

	for (size = i = 0; i < d->frames; i++) {
		struct bcm2835_dma_cb *control_block =
			&d->control_block_base[i];

		size += control_block->length;
	}

	return size;

Doing it this way, you avoid a bug here: your d->en is the period length
in bytes, and d->fn is the number of periods in the buffer.  Therefore,
the total size of the buffer is d->en * d->fn, and not
d->en * d->fn * d->es.

> +}
> +
> +static size_t bcm2835_dma_desc_size_pos(struct bcm2835_desc *d, dma_addr_t addr)
> +{
> +	unsigned i;
> +	size_t size;
> +
> +	for (size = i = 0; i < d->sglen; i++) {
> +		size_t this_size = bcm2835_dma_sg_size(&d->sg[i]);
> +
> +		if (size)
> +			size += this_size;
> +		else if (addr >= d->sg[i].addr &&
> +			 addr < d->sg[i].addr + this_size)
> +			size += d->sg[i].addr + this_size - addr;
> +	}

	for (size = i = 0; i < d->frames; i++) {
		struct bcm2835_dma_cb *control_block =
			&d->control_block_base[i];
		size_t this_size = control_block->length;
		dma_addr_t dma;

		if (d->dir == DMA_DEV_TO_MEM)
			dma = control_block->dst;
		else
			dma = control_block->src;

		if (size)
			size += this_size;
		else if (addr >= dma && addr < dma + this_size)
			size += dma + this_size - addr; 
	}

With these changes, you don't need d->sg[], d->sglen, d->es, d->en, or
d->fn.

> +	return size;
> +}
> +
> +
> +/*
> + * Returns current physical source address for the given DMA channel.
> + * If the channel is running the caller must disable interrupts prior calling
> + * this function and process the returned value before re-enabling interrupt to
> + * prevent races with the interrupt handler.

This comment could be better wrapped.

> + */
> +static dma_addr_t bcm2835_get_dma_src_pos(struct bcm2835_chan *c)
> +{
> +	return readl(c->dma_chan_base + BCM2835_DMA_SOURCE_AD);
> +}
> +
> +/*
> + * Returns current physical destination address for the given DMA channel.
> + * If the channel is running the caller must disable interrupts prior calling
> + * this function and process the returned value before re-enabling interrupt to
> + * prevent races with the interrupt handler.

As can this one.  Since the only caller is bcm2835_dma_tx_status(),
which disables interrupts, is this comment required?  Can we just move
the reading of these registers there?

> + */
> +static dma_addr_t bcm2835_get_dma_dst_pos(struct bcm2835_chan *c)
> +{
> +	return readl(c->dma_chan_base + BCM2835_DMA_DEST_AD);
> +}
> +
> +static enum dma_status bcm2835_dma_tx_status(struct dma_chan *chan,
> +	dma_cookie_t cookie, struct dma_tx_state *txstate)
> +{
> +	struct bcm2835_chan *c = to_bcm2835_dma_chan(chan);
> +	struct virt_dma_desc *vd;
> +	enum dma_status ret;
> +	unsigned long flags;
> +
> +	ret = dma_cookie_status(chan, cookie, txstate);
> +	if (ret == DMA_SUCCESS || !txstate)
> +		return ret;
> +
> +	spin_lock_irqsave(&c->vc.lock, flags);
> +	vd = vchan_find_desc(&c->vc, cookie);
> +	if (vd) {
> +		txstate->residue =
> +			bcm2835_dma_desc_size(to_bcm2835_dma_desc(&vd->tx));
> +	} else if (c->desc && c->desc->vd.tx.cookie == cookie) {
> +		struct bcm2835_desc *d = c->desc;
> +		dma_addr_t pos;
> +
> +		if (d->dir == DMA_MEM_TO_DEV)
> +			pos = bcm2835_get_dma_src_pos(c);
> +		else if (d->dir == DMA_DEV_TO_MEM)
> +			pos = bcm2835_get_dma_dst_pos(c);
> +		else
> +			pos = 0;
> +
> +		txstate->residue = bcm2835_dma_desc_size_pos(d, pos);
> +	} else {
> +		txstate->residue = 0;
> +	}
> +
> +	spin_unlock_irqrestore(&c->vc.lock, flags);
> +
> +	return ret;
> +}
> +
> +static void bcm2835_dma_issue_pending(struct dma_chan *chan)
> +{
> +	struct bcm2835_chan *c = to_bcm2835_dma_chan(chan);
> +	unsigned long flags;
> +
> +	spin_lock_irqsave(&c->vc.lock, flags);
> +	if (vchan_issue_pending(&c->vc) && !c->desc) {
> +		struct bcm2835_dmadev *d = to_bcm2835_dma_dev(chan->device);
> +		spin_lock(&d->lock);
> +		if (list_empty(&c->node))
> +			list_add_tail(&c->node, &d->pending);
> +		spin_unlock(&d->lock);
> +		tasklet_schedule(&d->task);
> +	}
> +	spin_unlock_irqrestore(&c->vc.lock, flags);
> +}
> +
> +
> +static struct dma_async_tx_descriptor *bcm2835_dma_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 bcm2835_chan *c = to_bcm2835_dma_chan(chan);
> +	enum dma_slave_buswidth dev_width;
> +	struct bcm2835_desc *d;
> +	dma_addr_t dev_addr;
> +	unsigned int es, sync_type;
> +	int frame;
> +	int frames;

I'd suggest unsigned for both frame and frames.  period_len and buf_len
are both unsigned, so period_len / buf_len will also be unsigned.

I much prefer unsigned for variables which are used as positive only
iterators and size-related stuff.  It helps to avoid the "oops I forgot
to check for negative values" problem.

> +
> +	/* Grab configuration */
> +	if (direction == DMA_DEV_TO_MEM) {
> +		dev_addr = c->cfg.src_addr;
> +		dev_width = c->cfg.src_addr_width;
> +		sync_type = BCM2835_DMA_S_DREQ;
> +	} else if (direction == DMA_MEM_TO_DEV) {
> +		dev_addr = c->cfg.dst_addr;
> +		dev_width = c->cfg.dst_addr_width;
> +		sync_type = BCM2835_DMA_D_DREQ;
> +	} else {
> +		dev_err(chan->device->dev, "%s: bad direction?\n", __func__);
> +		return NULL;
> +	}
> +
> +	/* Bus width translates to the element size (ES) */
> +	switch (dev_width) {
> +	case DMA_SLAVE_BUSWIDTH_4_BYTES:
> +		es = BCM2835_DMA_DATA_TYPE_S32;
> +		break;
> +	default:
> +		return NULL;
> +	}
> +
> +	/* Now allocate and setup the descriptor. */
> +	d = kzalloc(sizeof(*d) + sizeof(d->sg[0]), GFP_ATOMIC);
> +	if (!d)
> +		return NULL;

Obviously you won't need d->sg[0] here...

> +
> +	d->dir = direction;
> +	d->dev_addr = dev_addr;
> +	d->es = es;

You won't need this anymore, as you don't need to multiply by it above.

> +	d->sg[0].addr = buf_addr;
> +	d->sg[0].en = period_len;

You don't need to assign to d->sg[0].en with the changes below.

> +	d->sg[0].fn = buf_len / period_len;

Replace with:
	d->frames = buf_len / period_len;

> +	d->sglen = 1;

Don't need this.

> +
> +	/* Allocate memory for control blocks */
> +	d->control_block_size = d->sg[0].fn*sizeof(struct bcm2835_dma_cb);

Replace with:
	d->control_block_size = d->frames * sizeof(struct bcm2835_dma_cb);

And note the spaces around operators - that's part of the coding style.

> +	d->control_block_base = dma_alloc_coherent(chan->device->dev,
> +			d->control_block_size, &d->control_block_base_phys,
> +			GFP_KERNEL);
> +
> +	if (!d->control_block_base) {
> +		dev_err(chan->device->dev,
> +				"%s: Memory allocation error\n", __func__);
> +		return NULL;
> +	}
> +
> +	memset(d->control_block_base, 0, d->control_block_size);
> +
> +	/*
> +	 * Iterate over all frames, create a control block
> +	 * for each frame and link them together.
> +	 */
> +	frames = d->sg[0].fn;

Hence this is no longer required and can be removed - obviously
references to frames below will need to be d->frames.

> +
> +	for (frame = 0; frame < frames; frame++) {
> +		struct bcm2835_dma_cb *control_block =
> +			&d->control_block_base[frame];
> +
> +		/* Setup adresses */
> +		if (d->dir == DMA_DEV_TO_MEM) {
> +			control_block->info = BCM2835_DMA_D_INC;
> +			control_block->src = d->dev_addr;
> +			control_block->dst = d->sg[0].addr+frame*d->sg[0].en;

Replace with:
			control_block->dst = buf_addr + frame * period_len;

> +		} else {
> +			control_block->info = BCM2835_DMA_S_INC;
> +			control_block->src = d->sg[0].addr+frame*d->sg[0].en;

Replace with:
			control_block->src = buf_addr + frame * period_len;

> +			control_block->dst = d->dev_addr;
> +		}
> +
> +		/* Enable interrupt */
> +		control_block->info |= BCM2835_DMA_INT_EN;
> +
> +		/* Setup synchronization */
> +		if (sync_type != 0)
> +			control_block->info |= sync_type;
> +
> +		/* Setup DREQ channel */
> +		if (c->dreq != 0)
> +			control_block->info |=
> +				BCM2835_DMA_PER_MAP(c->dreq);
> +
> +		/* Length of a frame */
> +		control_block->length = d->sg[0].en;

Replace with:
		control_block->length = period_len;

> +
> +		/*
> +		 * Next block is the next frame.
> +		 * This DMA engine driver currently only supports cyclic DMA.
> +		 * Therefore, wrap around at number of frames.
> +		 */
> +		control_block->next = d->control_block_base_phys +
> +			sizeof(struct bcm2835_dma_cb)*((frame+1)%(frames));
> +
> +		/* The following fields are not used here */
> +		control_block->stride = 0;
> +		control_block->pad[0] = 0;
> +		control_block->pad[1] = 0;

Are the pad fields used by the hardware at all?  If not, you _could_
store the buffer address for this segment of the buffer here in native
endian format, and thereby simplify the bcm2835_dma_desc_size_pos()
too - you wouldn't then have to worry about the direction of the
transfer in bcm2835_dma_desc_size_pos() either.

> +	}
> +
> +	c->cyclic = true; /* nothing else is implemented */
> +
> +	return vchan_tx_prep(&c->vc, &d->vd, flags);
> +}
> +
> +static int bcm2835_dma_slave_config(struct bcm2835_chan *c,
> +		struct dma_slave_config *cfg)
> +{
> +	if ((cfg->direction == DMA_DEV_TO_MEM
> +			&& cfg->src_addr_width != DMA_SLAVE_BUSWIDTH_4_BYTES) ||
> +	    (cfg->direction == DMA_MEM_TO_DEV
> +			&& cfg->dst_addr_width != DMA_SLAVE_BUSWIDTH_4_BYTES)) {

Much better to put the && at the end of the preceding line.  Also, it
would be a good idea to align the "cfg->" parts below each other.

> +		return -EINVAL;
> +	}
> +
> +	memcpy(&c->cfg, cfg, sizeof(c->cfg));
> +
> +	return 0;
> +}
> +
> +static int bcm2835_dma_terminate_all(struct bcm2835_chan *c)
> +{
> +	struct bcm2835_dmadev *d = to_bcm2835_dma_dev(c->vc.chan.device);
> +	unsigned long flags;
> +	LIST_HEAD(head);
> +
> +	spin_lock_irqsave(&c->vc.lock, flags);
> +
> +	/* Prevent this channel being scheduled */
> +	spin_lock(&d->lock);
> +	list_del_init(&c->node);
> +	spin_unlock(&d->lock);
> +
> +	/*
> +	 * Stop DMA activity: we assume the callback will not be called
> +	 * after bcm_dma_abort() returns (even if it does, it will see
> +	 * c->desc is NULL and exit.)
> +	 */
> +	if (c->desc) {
> +		c->desc = NULL;
> +		bcm2835_dma_abort(c->dma_chan_base);
> +
> +		/* Wait for stopping */
> +		while (readl(c->dma_chan_base + BCM2835_DMA_CS)
> +			& BCM2835_DMA_ACTIVE)
> +			;

Hmm, how long can it take to stop?  At least adding cpu_relax() here
would be a good idea...

> +	}
> +
> +	vchan_get_all_descriptors(&c->vc, &head);
> +	spin_unlock_irqrestore(&c->vc.lock, flags);
> +	vchan_dma_desc_free_list(&c->vc, &head);
> +
> +	return 0;
> +}
> +
> +static int bcm2835_dma_pause(struct bcm2835_chan *c)
> +{
> +	/* FIXME: not supported by platform private API */
> +	return -EINVAL;
> +}
> +
> +static int bcm2835_dma_resume(struct bcm2835_chan *c)
> +{
> +	/* FIXME: not supported by platform private API */
> +	return -EINVAL;
> +}

Please remove these, and instead just assign -EINVAL directly below.

> +
> +static int bcm2835_dma_control(struct dma_chan *chan, enum dma_ctrl_cmd cmd,
> +	unsigned long arg)
> +{
> +	struct bcm2835_chan *c = to_bcm2835_dma_chan(chan);
> +	int ret;
> +
> +	switch (cmd) {
> +	case DMA_SLAVE_CONFIG:
> +		ret = bcm2835_dma_slave_config(c,
> +				(struct dma_slave_config *)arg);
> +		break;
> +
> +	case DMA_TERMINATE_ALL:
> +		ret = bcm2835_dma_terminate_all(c);
> +		break;
> +
> +	case DMA_PAUSE:
> +		ret = bcm2835_dma_pause(c);
> +		break;
> +
> +	case DMA_RESUME:
> +		ret = bcm2835_dma_resume(c);
> +		break;
> +
> +	default:
> +		ret = -ENXIO;
> +		break;
> +	}
> +
> +	return ret;
> +}
> +
> +static int bcm2835_dma_chan_init(struct bcm2835_dmadev *od, int dma_sig)
> +{
> +	struct bcm2835_chan *c;
> +
> +	c = kzalloc(sizeof(*c), GFP_KERNEL);
> +	if (!c)
> +		return -ENOMEM;
> +
> +	c->dma_sig = dma_sig;
> +	c->vc.desc_free = bcm2835_dma_desc_free;
> +	vchan_init(&c->vc, &od->ddev);
> +	INIT_LIST_HEAD(&c->node);
> +
> +	od->ddev.chancnt++;
> +
> +	return 0;
> +}
> +
> +static void bcm2835_dma_free(struct bcm2835_dmadev *od)
> +{
> +	tasklet_kill(&od->task);
> +	while (!list_empty(&od->ddev.channels)) {
> +		struct bcm2835_chan *c = list_first_entry(&od->ddev.channels,
> +			struct bcm2835_chan, vc.chan.device_node);
> +
> +		list_del(&c->vc.chan.device_node);
> +		tasklet_kill(&c->vc.task);
> +		kfree(c);
> +	}
> +}
> +
> +#if defined(CONFIG_OF)
> +static const struct of_device_id bcm2835_dma_of_match[] = {
> +	{
> +		.compatible = "brcm,bcm2835-dma",
> +	}
> +};
> +MODULE_DEVICE_TABLE(of, bcm2835_dma_of_match);
> +#endif
> +
> +static struct dma_chan *bcm2835_dma_xlate(struct of_phandle_args *dma_spec,
> +					   struct of_dma *ofdma)
> +{
> +	struct bcm2835_dmadev *d = ofdma->of_dma_data;
> +	struct dma_chan *chan, *candidate;
> +
> +retry:
> +	candidate = NULL;
> +
> +	/* walk the list of channels registered with the current instance and
> +	 * find one that is currently unused */
> +	list_for_each_entry(chan, &d->ddev.channels, device_node)
> +		if (chan->client_count == 0) {
> +			candidate = chan;
> +			break;
> +		}
> +
> +	if (!candidate)
> +		return NULL;
> +
> +	/* dma_get_slave_channel will return NULL if we lost a race between
> +	 * the lookup and the reservation */
> +	chan = dma_get_slave_channel(candidate);
> +
> +	if (chan) {
> +		struct bcm2835_chan *c = to_bcm2835_dma_chan(chan);
> +
> +		/* Set DREQ from param */
> +		c->dreq = dma_spec->args[0];
> +
> +		return chan;
> +	}
> +
> +	goto retry;
> +}
> +
> +static int bcm2835_dma_probe(struct platform_device *pdev)
> +{
> +	struct bcm2835_dmadev *od;
> +	struct resource *dma_res = NULL;
> +	void __iomem *dma_base = NULL;
> +	int rc = 0;
> +	int i;
> +	struct resource *irq;
> +	int irq_resources;

A good idea here would be:

	rc = dma_set_mask(&pdev->dev, DMA_BIT_MASK(32));
	if (rc)
		return rc;
	dma_set_coherent_mask(&pdev->dev, DMA_BIT_MASK(32));

since you can only handle 32-bit DMA - and the above is good practise to
do, even when the default mask is correct for the driver.  There's comments
in the DMA API documentation (under the Documentation/ directory) which
point at this.

One further comment on that is in this merge window, I'm pushing a series
of patches which simplify the above sequence, which will become:

	rc = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
	if (rc)
		return rc;

> +
> +	od = devm_kzalloc(&pdev->dev, sizeof(*od), GFP_KERNEL);
> +	if (!od)
> +		return -ENOMEM;
> +
> +	dma_res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
> +	dma_base = devm_ioremap_resource(&pdev->dev, dma_res);
> +	if (IS_ERR(dma_base))
> +		return PTR_ERR(dma_base);
> +
> +	od->dma_base = dma_base;
> +
> +	dma_cap_set(DMA_SLAVE, od->ddev.cap_mask);
> +	dma_cap_set(DMA_CYCLIC, od->ddev.cap_mask);
> +	od->ddev.device_alloc_chan_resources = bcm2835_dma_alloc_chan_resources;
> +	od->ddev.device_free_chan_resources = bcm2835_dma_free_chan_resources;
> +	od->ddev.device_tx_status = bcm2835_dma_tx_status;
> +	od->ddev.device_issue_pending = bcm2835_dma_issue_pending;
> +	od->ddev.device_prep_dma_cyclic = bcm2835_dma_prep_dma_cyclic;
> +	od->ddev.device_control = bcm2835_dma_control;
> +	od->ddev.dev = &pdev->dev;
> +	INIT_LIST_HEAD(&od->ddev.channels);
> +	INIT_LIST_HEAD(&od->pending);
> +	spin_lock_init(&od->lock);
> +
> +	tasklet_init(&od->task, bcm2835_dma_sched, (unsigned long)od);
> +
> +	irq_resources = 0;
> +
> +	for (i = 0; i < pdev->num_resources; i++) {
> +		if (IORESOURCE_IRQ == resource_type(&pdev->resource[i]))
> +			irq_resources++;

better would be (to remain compatible with your above code):
		if (platform_get_resource(pdev, IORESOURCE_IRQ, i) >= 0)
			irq_resources++;

or just use > 0.


> +	}
> +
> +	od->dma_irq_numbers = devm_kzalloc(&pdev->dev,
> +			irq_resources*sizeof(int), GFP_KERNEL);
> +	if (!od)
> +		return -ENOMEM;
> +
> +	for (i = 0; i < irq_resources; i++) {
> +		rc = bcm2835_dma_chan_init(od, i);
> +		if (rc) {
> +			bcm2835_dma_free(od);
> +			return rc;
> +		}
> +
> +		irq = platform_get_resource(pdev, IORESOURCE_IRQ, i);
> +		if (!irq) {
> +			dev_err(&pdev->dev,
> +					"No IRQ resource for channel %i\n", i);
> +			return -ENODEV;
> +		}
> +		od->dma_irq_numbers[i] = irq->start;
> +	}
> +
> +	rc = dma_async_device_register(&od->ddev);
> +	if (rc) {
> +		dev_err(&pdev->dev,
> +			"Failed to register slave DMA engine device: %d\n", rc);
> +		bcm2835_dma_free(od);
> +		return rc;
> +	}
> +
> +	platform_set_drvdata(pdev, od);
> +
> +	if (pdev->dev.of_node) {
> +		const void *chan_mask;
> +
> +		/* Device-tree DMA controller registration */
> +		rc = of_dma_controller_register(pdev->dev.of_node,
> +				bcm2835_dma_xlate, od);
> +		if (rc) {
> +			dev_err(&pdev->dev, "Failed to register DMA controller\n");
> +			dma_async_device_unregister(&od->ddev);
> +			bcm2835_dma_free(od);
> +			return rc;
> +		}
> +
> +		/* Request DMA channel mask from device tree */
> +		chan_mask = of_get_property(pdev->dev.of_node,
> +				"dma-channel-mask", NULL);
> +
> +		if (!chan_mask) {
> +			dev_err(&pdev->dev, "Failed to get channel mask\n");
> +			dma_async_device_unregister(&od->ddev);
> +			bcm2835_dma_free(od);
> +			return -EINVAL;
> +		}
> +
> +		od->chans_available = be32_to_cpup(chan_mask);
> +	}
> +
> +	dev_dbg(&pdev->dev, "Load BCM2835 DMA engine driver\n");
> +
> +	return rc;
> +}
> +
> +static int bcm2835_dma_remove(struct platform_device *pdev)
> +{
> +	struct bcm2835_dmadev *od = platform_get_drvdata(pdev);
> +
> +	dma_async_device_unregister(&od->ddev);
> +	bcm2835_dma_free(od);
> +
> +	return 0;
> +}
> +
> +static struct platform_driver bcm2835_dma_driver = {
> +	.probe	= bcm2835_dma_probe,
> +	.remove	= bcm2835_dma_remove,
> +	.driver = {
> +		.name = "bcm2835-dma",
> +		.owner = THIS_MODULE,
> +		.of_match_table = of_match_ptr(bcm2835_dma_of_match),
> +	},
> +};
> +
> +static const struct platform_device_info bcm2835_dma_dev_info = {
> +	.name = "bcm2835-dma",
> +	.id = -1,
> +	.dma_mask = DMA_BIT_MASK(32),
> +};

This doesn't appear to be referenced, can it be removed?

> +
> +static int bcm2835_dma_init(void)
> +{
> +	int rc = platform_driver_register(&bcm2835_dma_driver);
> +	return rc;
> +}
> +subsys_initcall(bcm2835_dma_init);
> +
> +static void __exit bcm2835_dma_exit(void)
> +{
> +	platform_driver_unregister(&bcm2835_dma_driver);
> +}
> +module_exit(bcm2835_dma_exit);

The above init/exit functions can just be:

module_platform_driver(&bcm2835_dma_driver);

That's it - thanks.  There's quite a number of comments in the above, and
I've made quite a number of suggestions there about reorganising the code
some more.  I think that is beneficial, because you don't need many of the
OMAP bits to handle this device, especially as the device uses control
blocks in memory.

The OMAP 'sg' stuff is only there because it doesn't use that facility,
and the driver has to manually program the DMA hardware for each segment
of the DMA.  You don't need that.