[PATCH] dmaengine: pl08x: mark slave channels as DMA_PRIVATE

[PATCH] dmaengine: pl08x: mark slave channels as DMA_PRIVATE

[Andreas Gabriel-Platschek]

The pl08x slave engine channels are only ever handed out through
explicit requests: DT clients go via pl08x_of_xlate() and the
platform-data path uses dma_request_channel() with a filter, both of
which end up in dma_get_slave_channel(). They are not meant to be
published in the general-purpose channel table.

Mark the slave device with DMA_PRIVATE at probe time so the dmaengine
core treats it as always-private. Without this, if a dmaengine client
is already active when the device is registered,
dma_async_device_register() takes a reference on every channel, and
the subsequent dma_get_slave_channel() from pl08x_of_xlate() fails
because the channel is no longer free (client_count != 0), leaving DT
slave clients unable to acquire channels.

The memcpy engine stays public as it is intended to be shared through
the per-CPU channel table.

Signed-off-by: Andreas Gabriel-Platschek <andi.platschek@gmail.com>
---
 drivers/dma/amba-pl08x.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/dma/amba-pl08x.c b/drivers/dma/amba-pl08x.c
index 543c2eb8cfa8..06881d3f29a9 100644
--- a/drivers/dma/amba-pl08x.c
+++ b/drivers/dma/amba-pl08x.c
@@ -2775,6 +2775,7 @@ static int pl08x_probe(struct amba_device *adev, const struct amba_id *id)
 		pl08x->has_slave = true;
 		dma_cap_set(DMA_SLAVE, pl08x->slave.cap_mask);
 		dma_cap_set(DMA_CYCLIC, pl08x->slave.cap_mask);
+		dma_cap_set(DMA_PRIVATE, pl08x->slave.cap_mask);
 		pl08x->slave.dev = &adev->dev;
 		pl08x->slave.device_free_chan_resources =
 			pl08x_free_chan_resources;
-- 
2.47.3

Re: [PATCH] dmaengine: pl08x: mark slave channels as DMA_PRIVATE

[Frank Li]

On Tue, Oct 27, 2026 at 08:00:37AM +0100, Andreas Gabriel-Platschek wrote:
> [You don't often get email from andi.platschek@gmail.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
>
> The pl08x slave engine channels are only ever handed out through
> explicit requests: DT clients go via pl08x_of_xlate() and the
> platform-data path uses dma_request_channel() with a filter, both of
> which end up in dma_get_slave_channel(). They are not meant to be
> published in the general-purpose channel table.
>
> Mark the slave device with DMA_PRIVATE at probe time so the dmaengine
> core treats it as always-private. Without this, if a dmaengine client
> is already active when the device is registered,
> dma_async_device_register() takes a reference on every channel,

Code change is okay. Are you sure the above description correct? where
code set it active when dma_async_device_register().

Frank

> and
> the subsequent dma_get_slave_channel() from pl08x_of_xlate() fails
> because the channel is no longer free (client_count != 0), leaving DT
> slave clients unable to acquire channels.
>
> The memcpy engine stays public as it is intended to be shared through
> the per-CPU channel table.
>
> Signed-off-by: Andreas Gabriel-Platschek <andi.platschek@gmail.com>
> ---
>  drivers/dma/amba-pl08x.c | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/drivers/dma/amba-pl08x.c b/drivers/dma/amba-pl08x.c
> index 543c2eb8cfa8..06881d3f29a9 100644
> --- a/drivers/dma/amba-pl08x.c
> +++ b/drivers/dma/amba-pl08x.c
> @@ -2775,6 +2775,7 @@ static int pl08x_probe(struct amba_device *adev, const struct amba_id *id)
>                 pl08x->has_slave = true;
>                 dma_cap_set(DMA_SLAVE, pl08x->slave.cap_mask);
>                 dma_cap_set(DMA_CYCLIC, pl08x->slave.cap_mask);
> +               dma_cap_set(DMA_PRIVATE, pl08x->slave.cap_mask);
>                 pl08x->slave.dev = &adev->dev;
>                 pl08x->slave.device_free_chan_resources =
>                         pl08x_free_chan_resources;
> --
> 2.47.3
>

Re: [PATCH] dmaengine: pl08x: mark slave channels as DMA_PRIVATE

[Andreas Gabriel-Platschek]

On 10/27/26 3:39 PM, Frank Li wrote:
> On Tue, Oct 27, 2026 at 08:00:37AM +0100, Andreas Gabriel-Platschek wrote:
>> [You don't often get email from andi.platschek@gmail.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
>>
>> The pl08x slave engine channels are only ever handed out through
>> explicit requests: DT clients go via pl08x_of_xlate() and the
>> platform-data path uses dma_request_channel() with a filter, both of
>> which end up in dma_get_slave_channel(). They are not meant to be
>> published in the general-purpose channel table.
>>
>> Mark the slave device with DMA_PRIVATE at probe time so the dmaengine
>> core treats it as always-private. Without this, if a dmaengine client
>> is already active when the device is registered,
>> dma_async_device_register() takes a reference on every channel,
> Code change is okay. Are you sure the above description correct? where
> code set it active when dma_async_device_register().
>
> Frank

I do think so, because

async_damengine_get() -> dmaengine_get() is called from 
crypto/async_tx/async_tx.c and increments the dmaengine_ref_count.
This way the dmaengine_ref_count is already >=1 before the pl08x 
dmaengine probes.

dma_async_device_register() (in dmaengine.c) iterates over the channels 
inside of the following conditional:

  if (dmaengine_ref_count && !dma_has_cap(DMA_PRIVATE, device->cap_mask))

and in there a dma_chan_get() is called for each channel and increments 
the client_count. With a client_count  >=1 we fail once 
pl08x_of_xlate()->dma_get_slave_channel() is called.

setting DMA_PRIVATE prevents the client_count from being incremented by 
the core, keeping the client_count at 0 until a real client requests one

hope that makes sense?
best!
Andi
>
>> and
>> the subsequent dma_get_slave_channel() from pl08x_of_xlate() fails
>> because the channel is no longer free (client_count != 0), leaving DT
>> slave clients unable to acquire channels.
>>
>> The memcpy engine stays public as it is intended to be shared through
>> the per-CPU channel table.
>>
>> Signed-off-by: Andreas Gabriel-Platschek <andi.platschek@gmail.com>
>> ---
>>   drivers/dma/amba-pl08x.c | 1 +
>>   1 file changed, 1 insertion(+)
>>
>> diff --git a/drivers/dma/amba-pl08x.c b/drivers/dma/amba-pl08x.c
>> index 543c2eb8cfa8..06881d3f29a9 100644
>> --- a/drivers/dma/amba-pl08x.c
>> +++ b/drivers/dma/amba-pl08x.c
>> @@ -2775,6 +2775,7 @@ static int pl08x_probe(struct amba_device *adev, const struct amba_id *id)
>>                  pl08x->has_slave = true;
>>                  dma_cap_set(DMA_SLAVE, pl08x->slave.cap_mask);
>>                  dma_cap_set(DMA_CYCLIC, pl08x->slave.cap_mask);
>> +               dma_cap_set(DMA_PRIVATE, pl08x->slave.cap_mask);
>>                  pl08x->slave.dev = &adev->dev;
>>                  pl08x->slave.device_free_chan_resources =
>>                          pl08x_free_chan_resources;
>> --
>> 2.47.3
>>

Re: [PATCH] dmaengine: pl08x: mark slave channels as DMA_PRIVATE

[Frank Li]

On Tue, Oct 27, 2026 at 07:25:10PM +0100, Andreas Gabriel-Platschek wrote:
> You don't often get email from andi.platschek@gmail.com. Learn why this is important<https://aka.ms/LearnAboutSenderIdentification>
>
>
> On Tue, Oct 27, 2026 at 3:39 PM Frank Li <Frank.li@oss.nxp.com<mailto:Frank.li@oss.nxp.com>> wrote:
> On Tue, Oct 27, 2026 at 08:00:37AM +0100, Andreas Gabriel-Platschek wrote:
> > [You don't often get email from andi.platschek@gmail.com<mailto:andi.platschek@gmail.com>. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
> >
> > The pl08x slave engine channels are only ever handed out through
> > explicit requests: DT clients go via pl08x_of_xlate() and the
> > platform-data path uses dma_request_channel() with a filter, both of
> > which end up in dma_get_slave_channel(). They are not meant to be
> > published in the general-purpose channel table.
> >
> > Mark the slave device with DMA_PRIVATE at probe time so the dmaengine
> > core treats it as always-private. Without this, if a dmaengine client
> > is already active when the device is registered,
> > dma_async_device_register() takes a reference on every channel,
>
> Code change is okay. Are you sure the above description correct? where
> code set it active when dma_async_device_register().
>
> I do think so, because
>
> async_damengine_get() -> dmaengine_get() is called from crypto/async_tx/async_tx.c and increments the dmaengine_ref_count.
> This way the dmaengine_ref_count is already >=1 before the pl08x dmaengine probes.
>
> dma_async_device_register() (in dmaengine.c) iterates over the channels inside of the following conditional:
>
>  if (dmaengine_ref_count && !dma_has_cap(DMA_PRIVATE, device->cap_mask))
>
> and in there a dma_chan_get() is called for each channel and increments the client_count. With a client_count  >=1 we fail once pl08x_of_xlate()->dma_get_slave_channel() is called.
>
> setting DMA_PRIVATE prevents the client_count from being incremented by the core, keeping the client_count at 0 until a real client requests one
>
> hope that makes sense?
> best!
> Andi

Reviewed-by: Frank Li <Frank.Li@nxp.com>

>
>
> Frank
>
> > and
> > the subsequent dma_get_slave_channel() from pl08x_of_xlate() fails
> > because the channel is no longer free (client_count != 0), leaving DT
> > slave clients unable to acquire channels.
> >
> > The memcpy engine stays public as it is intended to be shared through
> > the per-CPU channel table.
> >
> > Signed-off-by: Andreas Gabriel-Platschek <andi.platschek@gmail.com<mailto:andi.platschek@gmail.com>>
> > ---
> >  drivers/dma/amba-pl08x.c | 1 +
> >  1 file changed, 1 insertion(+)
> >
> > diff --git a/drivers/dma/amba-pl08x.c b/drivers/dma/amba-pl08x.c
> > index 543c2eb8cfa8..06881d3f29a9 100644
> > --- a/drivers/dma/amba-pl08x.c
> > +++ b/drivers/dma/amba-pl08x.c
> > @@ -2775,6 +2775,7 @@ static int pl08x_probe(struct amba_device *adev, const struct amba_id *id)
> >                 pl08x->has_slave = true;
> >                 dma_cap_set(DMA_SLAVE, pl08x->slave.cap_mask);
> >                 dma_cap_set(DMA_CYCLIC, pl08x->slave.cap_mask);
> > +               dma_cap_set(DMA_PRIVATE, pl08x->slave.cap_mask);
> >                 pl08x->slave.dev<http://slave.dev> = &adev->dev;
> >                 pl08x->slave.device_free_chan_resources =
> >                         pl08x_free_chan_resources;
> > --
> > 2.47.3
> >