Hi Rob,

On 30 August 2011 18:49, Rob Herring [off-list ref] wrote:

quoted

Thomas,

On 08/30/2011 07:18 AM, Thomas Abraham wrote:

quoted

Hi Rob,

On 26 August 2011 18:46, Rob Herring <robherring2@gmail.com
<mailto:robherring2@gmail.com>> wrote:

    Thomas,

    On 08/26/2011 03:40 AM, Thomas Abraham wrote:
    > For PL330 dma controllers instantiated from device tree, the channel
    > lookup is based on phandle of the dma controller and dma request id
    > specified by the client node. During probe, the private data of each
    > channel of the controller is set to point to the device node of the
    > dma controller. The 'chan_id' of the each channel is used as the
    > dma request id.
    >
    > Client driver requesting dma channels specify the phandle of the
    > dma controller and the request id. The pl330 filter function
    > converts the phandle to the device node pointer and matches that
    > with channel's private data. If a match is found, the request id
    > from the client node and the 'chan_id' of the channel is matched.
    > A channel is found if both the values match.
    >
    > Signed-off-by: Thomas Abraham <thomas.abraham@linaro.org
    <mailto:thomas.abraham@linaro.org>>
    > ---
    >  .../devicetree/bindings/dma/arm-pl330.txt          |   42
    +++++++++++++
    >  drivers/dma/pl330.c                                |   63
    +++++++++++++++++++-
    >  2 files changed, 103 insertions(+), 2 deletions(-)
    >  create mode 100644
    Documentation/devicetree/bindings/dma/arm-pl330.txt
    >
    > diff --git a/Documentation/devicetree/bindings/dma/arm-pl330.txt
    b/Documentation/devicetree/bindings/dma/arm-pl330.txt
    > new file mode 100644
    > index 0000000..46a8307
    > --- /dev/null
    > +++ b/Documentation/devicetree/bindings/dma/arm-pl330.txt
    > @@ -0,0 +1,42 @@
    > +* ARM PrimeCell PL330 DMA Controller
    > +
    > +The ARM PrimeCell PL330 DMA controller can move blocks of memory
    contents
    > +between memory and peripherals or memory to memory.
    > +
    > +Required properties:
    > +  - compatible: should one or more of the following
    > +    - arm,pl330-pdma - For controllers that support mem-to-dev
    and dev-to-mem
    > +      transfers.
    > +    - arm,pl330-mdma - For controllers that support mem-to-mem
    transfers only.

    And if they support both? I would think all controllers can support
    mem-to-mem. If so, the distinction can be made with the number of
    requests.


If a controller supports both types of transfer, the device node should
not claim compatibility for "arm,pl330-pdma" or "arm,pl330-mdma".
Compatible should be "arm,primecell".

No, every Primecell peripheral has "arm,primecell", so it is not
specific enough for a driver to use.

Is there a case that a controller with peripheral requests cannot
support mem-to-mem transfers? I don't think there is. You could decide
you don't want to for other reasons like you don't have enough free
channels, but that's really a s/w decision, not a h/w description.

Ok. The driver has now been changed in a way that DMA_MEMCPY
capability is assigned by default to a pl330 dma controller. The
DMA_SLAVE and DMA_CYCLIC capability are assigned if the controller
supports peripheral request interface. For mem-to-mem transfer channel
requests, the filter function specified for the request should check
for the right match.

quoted

quoted

    > +    - arm,primecell - should be included for all pl330 dma
    controller nodes.
    > +
    > +  - reg: physical base address of the controller and length of
    memory mapped
    > +    region.
    > +
    > +  - interrupts: interrupt number to the cpu.
    > +
    > +  - arm,primecell-periphid: should be 0x00041330.

    Should be optional. It's only needed when the h/w value is wrong. This
    is already documented in primecell.txt.


Ok. This will be made optional.



    > +
    > +  - arm,pl330-peri-reqs: number of actual peripheral requests
    connected to the
    > +    dma controller. Maximum value is 32.

    Perhaps could be a bitmask for sparsely populated requests. May not
    matter since phandles will define the connections.

    Can be optional and not present means 00 requests (mem-to-mem only).


As suggested by Russell, this property will be removed and its value
will be read from the configuration register.

Good. Reading a value of 0 requests can still be used to determine the
controller is mem-to-mem only.

quoted

    > +
    > +Example: (from Samsung's Exynos4 processor dtsi file)
    > +
    > +     pdma0: pdma@12680000 {
    > +             compatible = "arm,pl330-pdma", "arm,primecell";
    > +             reg = <0x12680000 0x1000>;
    > +             interrupts = <99>;
    > +             arm,primecell-periphid = <0x00041330>;
    > +             arm,pl330-peri-reqs = <30>;
    > +     };
    > +
    > +Client drivers (device nodes requiring dma transfers from
    dev-to-mem or
    > +mem-to-dev) should specify the DMA channel numbers using a
    two-value pair
    > +as shown below.
    > +
    > +  [property name]  = <[phandle of the dma controller] [dma
    request id]>;
    > +
    > +      where 'dma request id' is the dma request number which is
    connected
    > +      to the client controller.
    > +
    > +  Example:  tx-dma-channel = <&pdma0 12>;

    I like this approach. I looked at this some and some PPC platforms do a
    node for each channel/request, but this is much more simple and similar
    to clock binding approach.

    You need to define the property name. Probably just "dma-channel" is
    enough. For peripherals with more than 1, just list them out like when
    you have more than 1 interrupt. The order should be defined as part of
    that device's binding (i.e. 1st channel is tx and 2nd channel is rx).


I am little hesitant to do this the way you suggested. A controller
could have dma request lines connected to multiple dma controllers. So
the phandle could be different for each dma channel. Also, the client
drivers specify the property value for each dma channel requested (the
property value gets assigned to chan->private and then used by the
filter function to lookup the dma channel). So changing it the way you
have suggested would make things complex.

Perhaps you could make private point to the array entry rather than the
property.

But I don't have s strong opinion either way.

There could be cases where tx could be polled and rx uses dma transfer
and such this could vary across platforms. This would make parsing the
property value a little complex and so I still perfer to use property
for each dma channel required in the client device nodes.

quoted

quoted

    > diff --git a/drivers/dma/pl330.c b/drivers/dma/pl330.c
    > index 9732995..984dc18 100644
    > --- a/drivers/dma/pl330.c
    > +++ b/drivers/dma/pl330.c
    > @@ -19,6 +19,7 @@
    >  #include <linux/amba/pl330.h>
    >  #include <linux/pm_runtime.h>
    >  #include <linux/scatterlist.h>
    > +#include <linux/of.h>
    >
    >  #define NR_DEFAULT_DESC      16
    >
    > @@ -277,6 +278,20 @@ bool pl330_filter(struct dma_chan *chan, void
    *param)
    >       if (chan->device->dev->driver != &pl330_driver.drv)
    >               return false;
    >
    > +#ifdef CONFIG_OF
    > +     if (chan->device->dev->of_node) {
    > +             const __be32 *prop_value;
    > +             phandle phandle;
    > +             struct device_node *node;
    > +
    > +             prop_value = ((struct property *)param)->value;
    > +             phandle = be32_to_cpup(prop_value++);
    > +             node = of_find_node_by_phandle(phandle);
    > +             return ((chan->private == node) &&
    > +                             (chan->chan_id ==
    be32_to_cpup(prop_value)));
    > +     }
    > +#endif
    > +
    >       peri_id = chan->private;
    >       return *peri_id == (unsigned)param;
    >  }
    > @@ -777,6 +792,40 @@ static irqreturn_t pl330_irq_handler(int irq,
    void *data)
    >               return IRQ_NONE;
    >  }
    >
    > +#ifdef CONFIG_OF
    > +static struct dma_pl330_platdata *pl330_parse_dt(struct device *dev)
    > +{
    > +     struct dma_pl330_platdata *pdat;
    > +     const void *value;
    > +
    > +     pdat = devm_kzalloc(dev, sizeof(*pdat), GFP_KERNEL);
    > +     if (!pdat)
    > +             return NULL;

    Ideally, we will get rid of platform_data completely in the future, so I
    don't think filling it in from DT is the right approach.


Ok. I will drop the usage of platform data.


    > +
    > +     value = of_get_property(dev->of_node, "arm,pl330-peri-reqs",
    NULL);
    > +     if (value)
    > +             pdat->nr_valid_peri = be32_to_cpup(value);

    Can't you use the u32 helper function here?


This will go away now since the number of peripherals connected will be
read from the configuration register.


    > +
    > +     if (of_device_is_compatible(dev->of_node, "arm,pl330-pdma")) {
    > +             dma_cap_set(DMA_SLAVE, pdat->cap_mask);
    > +             dma_cap_set(DMA_CYCLIC, pdat->cap_mask);
    > +     } else if (of_device_is_compatible(dev->of_node,
    "arm,pl330-mdma")) {
    > +             dma_cap_set(DMA_MEMCPY, pdat->cap_mask);
    > +     } else if (of_device_is_compatible(dev->of_node,
    "arm,primecell")) {

    I don't think the driver should look at this property. This is really
    just for the bus code.


The dma capabilities are derived from the compatible value by the
driver. Sorry, I do not understand your suggestion for this.

"arm,primecell" is purely for identifying peripherals with the Primecell
ID registers and in turn used to create amba_device instances. You
cannot imply that it is a DMA controller.

This has been changed as you suggested. Could you have a look at the
updated patch listed below?


For PL330 dma controllers instantiated from device tree, the channel
lookup is based on phandle of the dma controller and dma request id
specified by the client node. During probe, the private data of each
channel of the controller is set to point to the device node of the
dma controller. The 'chan_id' of the each channel is used as the
dma request id.

Client driver requesting dma channels specify the phandle of the
dma controller and the request id. The pl330 filter function
converts the phandle to the device node pointer and matches that
with channel's private data. If a match is found, the request id
from the client node and the 'chan_id' of the channel is matched.
A channel is found if both the values match.

Signed-off-by: Thomas Abraham <redacted>
---
 .../devicetree/bindings/dma/arm-pl330.txt          |   29 ++++++++++++++++
 drivers/dma/pl330.c                                |   35 +++++++++++++++++---
 2 files changed, 59 insertions(+), 5 deletions(-)
 create mode 100644 Documentation/devicetree/bindings/dma/arm-pl330.txt

diff --git a/Documentation/devicetree/bindings/dma/arm-pl330.txt

b/Documentation/devicetree/bindings/dma/arm-pl330.txt
new file mode 100644
index 0000000..89f4b9c

--- /dev/null
+++ b/Documentation/devicetree/bindings/dma/arm-pl330.txt

@@ -0,0 +1,29 @@
+* ARM PrimeCell PL330 DMA Controller
+
+The ARM PrimeCell PL330 DMA controller can move blocks of memory contents
+between memory and peripherals or memory to memory.
+
+Required properties:
+  - compatible: should be "arm,primecell".

+
+      where 'dma request id' is the dma request number which is connected
+      to the client controller.
+
+  Example:  tx-dma-channel = <&pdma0 12>;

diff --git a/drivers/dma/pl330.c b/drivers/dma/pl330.c
index 9732995..0c55de4 100644
--- a/drivers/dma/pl330.c
+++ b/drivers/dma/pl330.c

@@ -19,6 +19,7 @@
 #include <linux/amba/pl330.h>
 #include <linux/pm_runtime.h>
 #include <linux/scatterlist.h>
+#include <linux/of.h>

 #define NR_DEFAULT_DESC	16

@@ -277,6 +278,20 @@ bool pl330_filter(struct dma_chan *chan, void *param)
 	if (chan->device->dev->driver != &pl330_driver.drv)
 		return false;

+#ifdef CONFIG_OF
+	if (chan->device->dev->of_node) {
+		const __be32 *prop_value;
+		phandle phandle;
+		struct device_node *node;
+
+		prop_value = ((struct property *)param)->value;
+		phandle = be32_to_cpup(prop_value++);
+		node = of_find_node_by_phandle(phandle);
+		return ((chan->private == node) &&
+				(chan->chan_id == be32_to_cpup(prop_value)));
+	}
+#endif
+
 	peri_id = chan->private;
 	return *peri_id == (unsigned)param;
 }

@@ -857,12 +872,17 @@ pl330_probe(struct amba_device *adev, const

struct amba_id *id)
 	INIT_LIST_HEAD(&pd->channels);

 	/* Initialize channel parameters */
-	num_chan = max(pdat ? pdat->nr_valid_peri : 0, (u8)pi->pcfg.num_chan);
+	num_chan = max(pdat ? pdat->nr_valid_peri : (u8)pi->pcfg.num_peri,
+			(u8)pi->pcfg.num_chan);
 	pdmac->peripherals = kzalloc(num_chan * sizeof(*pch), GFP_KERNEL);

 	for (i = 0; i < num_chan; i++) {
 		pch = &pdmac->peripherals[i];
-		pch->chan.private = pdat ? &pdat->peri_id[i] : NULL;
+		if (!adev->dev.of_node)
+			pch->chan.private = pdat ? &pdat->peri_id[i] : NULL;
+		else
+			pch->chan.private = adev->dev.of_node;
+
 		INIT_LIST_HEAD(&pch->work_list);
 		spin_lock_init(&pch->lock);
 		pch->pl330_chid = NULL;

@@ -876,11 +896,16 @@ pl330_probe(struct amba_device *adev, const

struct amba_id *id)
 	}

 	pd->dev = &adev->dev;
-	if (pdat)
+	if (pdat) {
 		pd->cap_mask = pdat->cap_mask;
-	else
+	} else {
 		dma_cap_set(DMA_MEMCPY, pd->cap_mask);
-
+		if (pi->pcfg.num_peri) {
+			dma_cap_set(DMA_SLAVE, pd->cap_mask);
+			dma_cap_set(DMA_CYCLIC, pd->cap_mask);
+		}	
+	}
+	
 	pd->device_alloc_chan_resources = pl330_alloc_chan_resources;
 	pd->device_free_chan_resources = pl330_free_chan_resources;
 	pd->device_prep_dma_memcpy = pl330_prep_dma_memcpy;