DMA client device driver usually needs to know at probe time whether
dma controller has been registered to deffer probe. So add a help
function of_dma_check_controller.
DMA request channel functions can also used to check it, but they
are usually called at open() time.
Signed-off-by: Richard Zhao <[email protected]>
---
drivers/dma/of-dma.c | 38 ++++++++++++++++++++++++++++++++++++++
include/linux/of_dma.h | 6 ++++++
2 files changed, 44 insertions(+)
diff --git a/drivers/dma/of-dma.c b/drivers/dma/of-dma.c
index e1c4d3b..b6828c1 100644
--- a/drivers/dma/of-dma.c
+++ b/drivers/dma/of-dma.c
@@ -188,6 +188,44 @@ struct dma_chan *of_dma_request_slave_channel(struct device_node *np,
}
/**
+ * of_dma_check_controller - Check whether dma controller registered
+ * @dev: pointer to client device structure
+ * @name: slave channel name
+ */
+int of_dma_check_controller(struct device *dev, const char *name)
+{
+ struct device_node *np = dev->of_node;
+ struct of_phandle_args dma_spec;
+ struct of_dma *ofdma = NULL;
+ int count, i;
+
+ if (!np || !name)
+ return -EINVAL;
+
+ count = of_property_count_strings(np, "dma-names");
+ if (count < 0) {
+ dev_err(dev, "dma-names property missing or empty\n");
+ return -EINVAL;
+ }
+
+ for (i = 0; i < count; i++) {
+ if (of_dma_match_channel(np, name, i, &dma_spec))
+ continue;
+
+ mutex_lock(&of_dma_lock);
+ ofdma = of_dma_find_controller(&dma_spec);
+ mutex_unlock(&of_dma_lock);
+ of_node_put(dma_spec.np);
+ }
+
+ if (ofdma)
+ return 0;
+ else
+ return -ENODEV;
+}
+EXPORT_SYMBOL_GPL(of_dma_check_controller);
+
+/**
* of_dma_simple_xlate - Simple DMA engine translation function
* @dma_spec: pointer to DMA specifier as found in the device tree
* @of_dma: pointer to DMA controller data
diff --git a/include/linux/of_dma.h b/include/linux/of_dma.h
index ae36298..bc7195f 100644
--- a/include/linux/of_dma.h
+++ b/include/linux/of_dma.h
@@ -39,6 +39,7 @@ extern int of_dma_controller_register(struct device_node *np,
extern void of_dma_controller_free(struct device_node *np);
extern struct dma_chan *of_dma_request_slave_channel(struct device_node *np,
const char *name);
+extern int of_dma_check_controller(struct device *dev, const char *name);
extern struct dma_chan *of_dma_simple_xlate(struct of_phandle_args *dma_spec,
struct of_dma *ofdma);
#else
@@ -60,6 +61,11 @@ static inline struct dma_chan *of_dma_request_slave_channel(struct device_node *
return NULL;
}
+static inline int of_dma_check_controller(struct device *dev, const char *name)
+{
+ return 0;
+}
+
static inline struct dma_chan *of_dma_simple_xlate(struct of_phandle_args *dma_spec,
struct of_dma *ofdma)
{
--
1.8.1.5
On 08/02/2013 12:04 AM, Richard Zhao wrote:
> DMA client device driver usually needs to know at probe time whether
> dma controller has been registered to deffer probe. So add a help
> function of_dma_check_controller.
>
> DMA request channel functions can also used to check it, but they
> are usually called at open() time.
> diff --git a/drivers/dma/of-dma.c b/drivers/dma/of-dma.c
> index e1c4d3b..b6828c1 100644
> +int of_dma_check_controller(struct device *dev, const char *name)
> + for (i = 0; i < count; i++) {
> + if (of_dma_match_channel(np, name, i, &dma_spec))
> + continue;
> +
> + mutex_lock(&of_dma_lock);
> + ofdma = of_dma_find_controller(&dma_spec);
> + mutex_unlock(&of_dma_lock);
> + of_node_put(dma_spec.np);
Do we need to add the following here:
if (ofdma)
break
To ensure that as soon as a successful match is found, the loop exits?
Otherwise, if there are multiple providers for that name, and the first
N are registered but the last isn't, this function will still return
failure.
> + if (ofdma)
> + return 0;
> + else
> + return -ENODEV;
That probably should be -EPROBE_DEFER?
Although, what about differentiating between "entry not found by
of_dma_match_channel" and "controller not yet probed"?
On Sat, Aug 03, 2013 at 03:59:59AM +0800, Stephen Warren wrote:
> On 08/02/2013 12:04 AM, Richard Zhao wrote:
> > DMA client device driver usually needs to know at probe time whether
> > dma controller has been registered to deffer probe. So add a help
> > function of_dma_check_controller.
> >
> > DMA request channel functions can also used to check it, but they
> > are usually called at open() time.
>
> > diff --git a/drivers/dma/of-dma.c b/drivers/dma/of-dma.c
> > index e1c4d3b..b6828c1 100644
>
> > +int of_dma_check_controller(struct device *dev, const char *name)
>
>
> > + for (i = 0; i < count; i++) {
> > + if (of_dma_match_channel(np, name, i, &dma_spec))
> > + continue;
> > +
> > + mutex_lock(&of_dma_lock);
> > + ofdma = of_dma_find_controller(&dma_spec);
> > + mutex_unlock(&of_dma_lock);
> > + of_node_put(dma_spec.np);
>
> Do we need to add the following here:
>
> if (ofdma)
> break
>
> To ensure that as soon as a successful match is found, the loop exits?
> Otherwise, if there are multiple providers for that name, and the first
> N are registered but the last isn't, this function will still return
> failure.
Right. Thanks.
>
> > + if (ofdma)
> > + return 0;
> > + else
> > + return -ENODEV;
>
> That probably should be -EPROBE_DEFER?
>
> Although, what about differentiating between "entry not found by
> of_dma_match_channel" and "controller not yet probed"?
I thought it might be called in non-probe functions. If no people
against it, I'll change it to EPROBE_DEFER.
Thanks
Richard
DMA client device driver usually needs to know at probe time whether
dma controller has been registered to deffer probe. So add a help
function of_dma_check_controller.
DMA request channel functions can also used to check it, but they
are usually called at open() time.
Signed-off-by: Richard Zhao <[email protected]>
---
drivers/dma/of-dma.c | 39 +++++++++++++++++++++++++++++++++++++++
include/linux/of_dma.h | 6 ++++++
2 files changed, 45 insertions(+)
diff --git a/drivers/dma/of-dma.c b/drivers/dma/of-dma.c
index 8cd5f3f..0063ddf 100644
--- a/drivers/dma/of-dma.c
+++ b/drivers/dma/of-dma.c
@@ -189,6 +189,45 @@ struct dma_chan *of_dma_request_slave_channel(struct device_node *np,
}
/**
+ * of_dma_check_controller - Check whether dma controller registered
+ * @dev: pointer to client device structure
+ * @name: slave channel name
+ */
+int of_dma_check_controller(struct device *dev, const char *name)
+{
+ struct device_node *np = dev->of_node;
+ struct of_phandle_args dma_spec;
+ struct of_dma *ofdma = NULL;
+ int count, i, ret = 0;
+
+ if (!np || !name)
+ return -EINVAL;
+
+ count = of_property_count_strings(np, "dma-names");
+ if (count < 0) {
+ dev_err(dev, "dma-names property missing or empty\n");
+ return -EINVAL;
+ }
+
+ for (i = 0; i < count; i++) {
+ ret = of_dma_match_channel(np, name, i, &dma_spec);
+ if (ret)
+ continue;
+
+ mutex_lock(&of_dma_lock);
+ ofdma = of_dma_find_controller(&dma_spec);
+ mutex_unlock(&of_dma_lock);
+ of_node_put(dma_spec.np);
+ if (!ofdma)
+ ret = -EPROBE_DEFER;
+ break;
+ }
+
+ return ret;
+}
+EXPORT_SYMBOL_GPL(of_dma_check_controller);
+
+/**
* of_dma_simple_xlate - Simple DMA engine translation function
* @dma_spec: pointer to DMA specifier as found in the device tree
* @of_dma: pointer to DMA controller data
diff --git a/include/linux/of_dma.h b/include/linux/of_dma.h
index ae36298..bc7195f 100644
--- a/include/linux/of_dma.h
+++ b/include/linux/of_dma.h
@@ -39,6 +39,7 @@ extern int of_dma_controller_register(struct device_node *np,
extern void of_dma_controller_free(struct device_node *np);
extern struct dma_chan *of_dma_request_slave_channel(struct device_node *np,
const char *name);
+extern int of_dma_check_controller(struct device *dev, const char *name);
extern struct dma_chan *of_dma_simple_xlate(struct of_phandle_args *dma_spec,
struct of_dma *ofdma);
#else
@@ -60,6 +61,11 @@ static inline struct dma_chan *of_dma_request_slave_channel(struct device_node *
return NULL;
}
+static inline int of_dma_check_controller(struct device *dev, const char *name)
+{
+ return 0;
+}
+
static inline struct dma_chan *of_dma_simple_xlate(struct of_phandle_args *dma_spec,
struct of_dma *ofdma)
{
--
1.8.1.5
On 08/22/2013 12:43 AM, Richard Zhao wrote:
> DMA client device driver usually needs to know at probe time whether
> dma controller has been registered to deffer probe. So add a help
> function of_dma_check_controller.
>
> DMA request channel functions can also used to check it, but they
> are usually called at open() time.
This new function is almost identical to the existing
of_dma_request_slave_channel(). Surely the code should be shared?
But that said, I don't see any need for a new function; why can't
drivers simply call of_dma_request_slave_channel() at probe time; from
what I can see, that function doesn't actually request the channel, but
rather simply looks it up, just like this one. The only difference is
that of_dma_xlate() is also called, but that's just doing some data
transformation, not actually recording channel ownership.
On Fri, Aug 23, 2013 at 04:36:53AM +0800, Stephen Warren wrote:
> On 08/22/2013 12:43 AM, Richard Zhao wrote:
> > DMA client device driver usually needs to know at probe time whether
> > dma controller has been registered to deffer probe. So add a help
> > function of_dma_check_controller.
> >
> > DMA request channel functions can also used to check it, but they
> > are usually called at open() time.
>
> This new function is almost identical to the existing
> of_dma_request_slave_channel(). Surely the code should be shared?
ofdma->of_dma_xlate(&dma_spec, ofdma);
The above is called holding of_dma_lock. If I want to abstract the
common lines, there' two options.
Option 1:
static struct of_dma* of_dma_check_controller_locked(np, name)
{
parameter check
get dma-names count and check return value
for loop to get of_dma
return PTR_ERR(err) or of_dma
}
struct dma_chan *of_dma_request_slave_channel(struct device_node *np,
const char *name)
{
chan = null;
mutex_lock(&of_dma_lock);
of_dma = of_dma_check_controller_locked(np, name)
if(!IS_ERR(of_dma))
chan = ofdma->of_dma_xlate(&dma_spec, ofdma);
mutex_unlock(&of_dma_lock);
return chan;
}
int of_dma_check_controller(struct device *dev, const char *name)
{
mutex_lock(&of_dma_lock);
ofdma = of_dma_check_controller_locked(dev->of_node, name);
mutex_unlock(&of_dma_lock);
if (IS_ERR(ofdma))
return ERR_PTR(ofdma);
else
return 0;
}
Option 2:
static struct of_dma* of_dma_check_controller_getlock(np, name)
{
parameter check
get dma-names count and check return value
for loop to get of_dma, get lock at old place
if failed, unlock.
return PTR_ERR(err) or of_dma
}
struct dma_chan *of_dma_request_slave_channel(struct device_node *np,
const char *name)
{
} of_dma = of_dma_check_controller_getlock(np, name)
if(!IS_ERR(of_dma)) {
chan = ofdma->of_dma_xlate(&dma_spec, ofdma);
unlock;
}
return chan;
}
int of_dma_check_controller(struct device *dev, const char *name)
ofdma = of_dma_check_controller_locked(dev->of_node, name);
if (IS_ERR(ofdma)) {
return ERR_PTR(ofdma);
} else {
unlock;
return 0;
}
}
> But that said, I don't see any need for a new function; why can't
> drivers simply call of_dma_request_slave_channel() at probe time;
It'll mislead user. channel supposed to be request at open time.
> from
> what I can see, that function doesn't actually request the channel, but
> rather simply looks it up, just like this one. The only difference is
> that of_dma_xlate() is also called, but that's just doing some data
> transformation, not actually recording channel ownership.
xlate function request the channel if things go well.
Thanks
Richard
On 08/22/2013 07:17 PM, Richard Zhao wrote:
> On Fri, Aug 23, 2013 at 04:36:53AM +0800, Stephen Warren wrote:
>> On 08/22/2013 12:43 AM, Richard Zhao wrote:
>>> DMA client device driver usually needs to know at probe time whether
>>> dma controller has been registered to deffer probe. So add a help
>>> function of_dma_check_controller.
>>>
>>> DMA request channel functions can also used to check it, but they
>>> are usually called at open() time.
>>
>> This new function is almost identical to the existing
>> of_dma_request_slave_channel(). Surely the code should be shared?
>
> ofdma->of_dma_xlate(&dma_spec, ofdma);
> The above is called holding of_dma_lock. If I want to abstract the
> common lines, there' two options.
What is the problem with acquiring the lock? If request_slave_channel()
needs to take the lock, then there must be a reason which presumably
applies to the other path too.
...
>> But that said, I don't see any need for a new function; why can't
>> drivers simply call of_dma_request_slave_channel() at probe time;
>
> It'll mislead user. channel supposed to be request at open time.
I don't agree.
>> from
>> what I can see, that function doesn't actually request the channel, but
>> rather simply looks it up, just like this one. The only difference is
>> that of_dma_xlate() is also called, but that's just doing some data
>> transformation, not actually recording channel ownership.
>
> xlate function request the channel if things go well.
Oh. xlate should not do that; that's a design flaw. xlate should do
nothing more than translate the DT content to an internal channel ID.