In etm_setup_aux(), when a user sink is obtained via
coresight_get_sink_by_id(), it increments the reference count of the
sink device. However, if the sink is used in path building, the path
holds a reference, but the initial reference from
coresight_get_sink_by_id() is not released, causing a reference count
leak. We should release the initial reference after the path is built.

Found by code review.

Cc: stable@vger.kernel.org
Fixes: 0e6c20517596 ("coresight: etm-perf: Allow an event to use different sinks")
Signed-off-by: Ma Ke <make24@iscas.ac.cn>
---
Changes in v2:
- modified the patch as suggestions.
---
 drivers/hwtracing/coresight/coresight-etm-perf.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.c b/drivers/hwtracing/coresight/coresight-etm-perf.c
index 17afa0f4cdee..56d012ab6d3a 100644
--- a/drivers/hwtracing/coresight/coresight-etm-perf.c
+++ b/drivers/hwtracing/coresight/coresight-etm-perf.c
@@ -454,6 +454,11 @@ static void *etm_setup_aux(struct perf_event *event, void **pages,
 		goto err;
 
 out:
+	if (user_sink) {
+		put_device(&user_sink->dev);
+		user_sink = NULL;
+	}
+
 	return event_data;
 
 err:
-- 
2.17.1

On 15/12/2025 04:27, Ma Ke wrote:
> In etm_setup_aux(), when a user sink is obtained via
> coresight_get_sink_by_id(), it increments the reference count of the
> sink device. However, if the sink is used in path building, the path
> holds a reference, but the initial reference from
> coresight_get_sink_by_id() is not released, causing a reference count
> leak. We should release the initial reference after the path is built.
> 
> Found by code review.
> 
> Cc: stable@vger.kernel.org
> Fixes: 0e6c20517596 ("coresight: etm-perf: Allow an event to use different sinks")
> Signed-off-by: Ma Ke <make24@iscas.ac.cn>
> ---
> Changes in v2:
> - modified the patch as suggestions.

I think Leo's comment on the previous v2 is still unaddressed. But 
releasing it in coresight_get_sink_by_id() would make it consistent with 
coresight_find_csdev_by_fwnode() and prevent further mistakes.

It also leads me to see that users of coresight_find_device_by_fwnode() 
should also release it, but only one out of two appears to.

James

> ---
>   drivers/hwtracing/coresight/coresight-etm-perf.c | 5 +++++
>   1 file changed, 5 insertions(+)
> 
> diff --git a/drivers/hwtracing/coresight/coresight-etm-perf.c b/drivers/hwtracing/coresight/coresight-etm-perf.c
> index 17afa0f4cdee..56d012ab6d3a 100644
> --- a/drivers/hwtracing/coresight/coresight-etm-perf.c
> +++ b/drivers/hwtracing/coresight/coresight-etm-perf.c
> @@ -454,6 +454,11 @@ static void *etm_setup_aux(struct perf_event *event, void **pages,
>   		goto err;
>   
>   out:
> +	if (user_sink) {
> +		put_device(&user_sink->dev);
> +		user_sink = NULL;
> +	}
> +
>   	return event_data;
>   
>   err:

On Mon, Dec 15, 2025 at 11:02:08AM +0200, James Clark wrote:
> 
> 
> On 15/12/2025 04:27, Ma Ke wrote:
> > In etm_setup_aux(), when a user sink is obtained via
> > coresight_get_sink_by_id(), it increments the reference count of the
> > sink device. However, if the sink is used in path building, the path
> > holds a reference, but the initial reference from
> > coresight_get_sink_by_id() is not released, causing a reference count
> > leak. We should release the initial reference after the path is built.
> > 
> > Found by code review.
> > 
> > Cc: stable@vger.kernel.org
> > Fixes: 0e6c20517596 ("coresight: etm-perf: Allow an event to use different sinks")
> > Signed-off-by: Ma Ke <make24@iscas.ac.cn>
> > ---
> > Changes in v2:
> > - modified the patch as suggestions.
> 
> I think Leo's comment on the previous v2 is still unaddressed. But releasing
> it in coresight_get_sink_by_id() would make it consistent with
> coresight_find_csdev_by_fwnode() and prevent further mistakes.

The point is the coresight core layer uses coresight_grab_device() to
increase the device's refcnt.  This is why we don't need to grab a
device when setup AUX.

> It also leads me to see that users of coresight_find_device_by_fwnode()
> should also release it, but only one out of two appears to.

Good finding!

Thanks,
Leo

On 12/15/2025 5:51 PM, Leo Yan wrote:
> On Mon, Dec 15, 2025 at 11:02:08AM +0200, James Clark wrote:
>>
>>
>> On 15/12/2025 04:27, Ma Ke wrote:
>>> In etm_setup_aux(), when a user sink is obtained via
>>> coresight_get_sink_by_id(), it increments the reference count of the
>>> sink device. However, if the sink is used in path building, the path
>>> holds a reference, but the initial reference from
>>> coresight_get_sink_by_id() is not released, causing a reference count
>>> leak. We should release the initial reference after the path is built.
>>>
>>> Found by code review.
>>>
>>> Cc: stable@vger.kernel.org
>>> Fixes: 0e6c20517596 ("coresight: etm-perf: Allow an event to use different sinks")
>>> Signed-off-by: Ma Ke <make24@iscas.ac.cn>
>>> ---
>>> Changes in v2:
>>> - modified the patch as suggestions.
>>
>> I think Leo's comment on the previous v2 is still unaddressed. But releasing
>> it in coresight_get_sink_by_id() would make it consistent with
>> coresight_find_csdev_by_fwnode() and prevent further mistakes.
> 
> The point is the coresight core layer uses coresight_grab_device() to
> increase the device's refcnt.  This is why we don't need to grab a
> device when setup AUX.

That make sense. We dont need to hold the refcnt for a while and it 
should be released immediately after locating the required device.

Thanks,
Jie

> 
>> It also leads me to see that users of coresight_find_device_by_fwnode()
>> should also release it, but only one out of two appears to.
> 
> Good finding!
> 
> Thanks,
> Leo
>

On 12/15/2025 10:09 AM, Jie Gan wrote:
> On 12/15/2025 5:51 PM, Leo Yan wrote:
>> On Mon, Dec 15, 2025 at 11:02:08AM +0200, James Clark wrote:
>>>
>>>
>>> On 15/12/2025 04:27, Ma Ke wrote:
>>>> In etm_setup_aux(), when a user sink is obtained via
>>>> coresight_get_sink_by_id(), it increments the reference count of the
>>>> sink device. However, if the sink is used in path building, the path
>>>> holds a reference, but the initial reference from
>>>> coresight_get_sink_by_id() is not released, causing a reference count
>>>> leak. We should release the initial reference after the path is built.
>>>>
>>>> Found by code review.
>>>>
>>>> Cc: stable@vger.kernel.org
>>>> Fixes: 0e6c20517596 ("coresight: etm-perf: Allow an event to use different sinks")
>>>> Signed-off-by: Ma Ke <make24@iscas.ac.cn>
>>>> ---
>>>> Changes in v2:
>>>> - modified the patch as suggestions.
>>>
>>> I think Leo's comment on the previous v2 is still unaddressed. But releasing
>>> it in coresight_get_sink_by_id() would make it consistent with
>>> coresight_find_csdev_by_fwnode() and prevent further mistakes.
>> 
>> The point is the coresight core layer uses coresight_grab_device() to
>> increase the device's refcnt.  This is why we don't need to grab a
>> device when setup AUX.
>
> That make sense. We dont need to hold the refcnt for a while and it 
> should be released immediately after locating the required device.
>
> Thanks,
> Jie

>> 
>>> It also leads me to see that users of coresight_find_device_by_fwnode()
>>> should also release it, but only one out of two appears to.
>> 
>> Good finding!
>> 
>> Thanks,
>> Leo
>> 
Hi all,

Thank you for the insightful discussion. I've carefully read the 
feedback from Leo, James, and Jie, and now have a clear understanding 
of the reference count management.

The core issue: coresight_get_sink_by_id() internally calls 
bus_find_device(), which increases reference count via get_device().

From the discussion, I note two possible fix directions:

1. Release the initial reference in etm_setup_aux() (current v2 patch)
2. Modify the behavior of coresight_get_sink_by_id() itself so it 
doesn't increase the reference count. 

Leo mentioned referencing how acpi_dev_present() does it, and James 
also pointed out that APIs should be consistent. I think it makes 
sense that following the principle like "lookup doesn't hold a 
reference" could prevent similar leaks in the future.

To ensure the correctness of the v3 patch, I'd like to confirm which 
patch is preferred. If option 2 is the consensus, I'm happy to modify 
the implementation of coresight_get_sink_by_id() as suggested.

Looking forward to your further guidance.

Thanks!

Hi,

On Fri, Dec 19, 2025 at 10:39:49AM +0800, Ma Ke wrote:

[...]

> From the discussion, I note two possible fix directions:
> 
> 1. Release the initial reference in etm_setup_aux() (current v2 patch)
> 2. Modify the behavior of coresight_get_sink_by_id() itself so it 
> doesn't increase the reference count. 

The option 2 is the right way to go.

> To ensure the correctness of the v3 patch, I'd like to confirm which 
> patch is preferred. If option 2 is the consensus, I'm happy to modify 
> the implementation of coresight_get_sink_by_id() as suggested.

It is good to use a separate patch to fix
coresight_find_device_by_fwnode() mentioned by James:

diff --git a/drivers/hwtracing/coresight/coresight-platform.c b/drivers/hwtracing/coresight/coresight-platform.c
index 0db64c5f4995..2b34f818ba88 100644
--- a/drivers/hwtracing/coresight/coresight-platform.c
+++ b/drivers/hwtracing/coresight/coresight-platform.c
@@ -107,14 +107,16 @@ coresight_find_device_by_fwnode(struct fwnode_handle *fwnode)
 	 * platform bus.
 	 */
 	dev = bus_find_device_by_fwnode(&platform_bus_type, fwnode);
-	if (dev)
-		return dev;
 
 	/*
 	 * We have a configurable component - circle through the AMBA bus
 	 * looking for the device that matches the endpoint node.
 	 */
-	return bus_find_device_by_fwnode(&amba_bustype, fwnode);
+	if (!dev)
+		dev = bus_find_device_by_fwnode(&amba_bustype, fwnode);
+
+	put_device(dev);
+	return dev;
 }
 
 /*
@@ -274,7 +276,6 @@ static int of_coresight_parse_endpoint(struct device *dev,
 
 	of_node_put(rparent);
 	of_node_put(rep);
-	put_device(rdev);
 
 	return ret;
 }

Thanks for working on this.

On 19/12/2025 09:41, Leo Yan wrote:
> Hi,
> 
> On Fri, Dec 19, 2025 at 10:39:49AM +0800, Ma Ke wrote:
> 
> [...]
> 
>>  From the discussion, I note two possible fix directions:
>>
>> 1. Release the initial reference in etm_setup_aux() (current v2 patch)
>> 2. Modify the behavior of coresight_get_sink_by_id() itself so it
>> doesn't increase the reference count.
> 
> The option 2 is the right way to go.
> 
>> To ensure the correctness of the v3 patch, I'd like to confirm which
>> patch is preferred. If option 2 is the consensus, I'm happy to modify
>> the implementation of coresight_get_sink_by_id() as suggested.
> 
> It is good to use a separate patch to fix
> coresight_find_device_by_fwnode() mentioned by James:
> 
> diff --git a/drivers/hwtracing/coresight/coresight-platform.c b/drivers/hwtracing/coresight/coresight-platform.c
> index 0db64c5f4995..2b34f818ba88 100644
> --- a/drivers/hwtracing/coresight/coresight-platform.c
> +++ b/drivers/hwtracing/coresight/coresight-platform.c
> @@ -107,14 +107,16 @@ coresight_find_device_by_fwnode(struct fwnode_handle *fwnode)
>   	 * platform bus.
>   	 */
>   	dev = bus_find_device_by_fwnode(&platform_bus_type, fwnode);
> -	if (dev)
> -		return dev;
>   
>   	/*
>   	 * We have a configurable component - circle through the AMBA bus
>   	 * looking for the device that matches the endpoint node.
>   	 */
> -	return bus_find_device_by_fwnode(&amba_bustype, fwnode);
> +	if (!dev)
> +		dev = bus_find_device_by_fwnode(&amba_bustype, fwnode);
> +
> +	put_device(dev);

^^ NAK, see below.

> +	return dev;
>   }
>   
>   /*
> @@ -274,7 +276,6 @@ static int of_coresight_parse_endpoint(struct device *dev,
>   
>   	of_node_put(rparent);
>   	of_node_put(rep);
> -	put_device(rdev);

This doesn't look good. We can't use the "dev" reliably without the 
reference count. We are opening up use-after-free.

NAK for this.

Suzuki

>   
>   	return ret;
>   }
> 
> Thanks for working on this.

On Fri, Dec 19, 2025 at 09:59:54AM +0000, Suzuki K Poulose wrote:

[...]

> > diff --git a/drivers/hwtracing/coresight/coresight-platform.c b/drivers/hwtracing/coresight/coresight-platform.c
> > index 0db64c5f4995..2b34f818ba88 100644
> > --- a/drivers/hwtracing/coresight/coresight-platform.c
> > +++ b/drivers/hwtracing/coresight/coresight-platform.c
> > @@ -107,14 +107,16 @@ coresight_find_device_by_fwnode(struct fwnode_handle *fwnode)
> >   	 * platform bus.
> >   	 */
> >   	dev = bus_find_device_by_fwnode(&platform_bus_type, fwnode);
> > -	if (dev)
> > -		return dev;
> >   	/*
> >   	 * We have a configurable component - circle through the AMBA bus
> >   	 * looking for the device that matches the endpoint node.
> >   	 */
> > -	return bus_find_device_by_fwnode(&amba_bustype, fwnode);
> > +	if (!dev)
> > +		dev = bus_find_device_by_fwnode(&amba_bustype, fwnode);
> > +
> > +	put_device(dev);
> 
> ^^ NAK, see below.
> 
> > +	return dev;
> >   }
> >   /*
> > @@ -274,7 +276,6 @@ static int of_coresight_parse_endpoint(struct device *dev,
> >   	of_node_put(rparent);
> >   	of_node_put(rep);
> > -	put_device(rdev);
> 
> This doesn't look good. We can't use the "dev" reliably without the
> reference count. We are opening up use-after-free.

My understanding is we don't grab a device from
coresight_find_device_by_fwnode().  The callers only check whether the
device is present on the bus; if it isn't, the driver defers probe.

This is similiar to coresight_find_csdev_by_fwnode(), which calls
put_device(dev) to release refcnt immediately.  This is why I
suggested the change, so the two functions behave consistently.

Thanks,
Leo

On 19/12/2025 11:38, Leo Yan wrote:
> On Fri, Dec 19, 2025 at 09:59:54AM +0000, Suzuki K Poulose wrote:
> 
> [...]
> 
>>> diff --git a/drivers/hwtracing/coresight/coresight-platform.c b/drivers/hwtracing/coresight/coresight-platform.c
>>> index 0db64c5f4995..2b34f818ba88 100644
>>> --- a/drivers/hwtracing/coresight/coresight-platform.c
>>> +++ b/drivers/hwtracing/coresight/coresight-platform.c
>>> @@ -107,14 +107,16 @@ coresight_find_device_by_fwnode(struct fwnode_handle *fwnode)
>>>    	 * platform bus.
>>>    	 */
>>>    	dev = bus_find_device_by_fwnode(&platform_bus_type, fwnode);
>>> -	if (dev)
>>> -		return dev;
>>>    	/*
>>>    	 * We have a configurable component - circle through the AMBA bus
>>>    	 * looking for the device that matches the endpoint node.
>>>    	 */
>>> -	return bus_find_device_by_fwnode(&amba_bustype, fwnode);
>>> +	if (!dev)
>>> +		dev = bus_find_device_by_fwnode(&amba_bustype, fwnode);
>>> +
>>> +	put_device(dev);
>>
>> ^^ NAK, see below.
>>
>>> +	return dev;
>>>    }
>>>    /*
>>> @@ -274,7 +276,6 @@ static int of_coresight_parse_endpoint(struct device *dev,
>>>    	of_node_put(rparent);
>>>    	of_node_put(rep);
>>> -	put_device(rdev);
>>
>> This doesn't look good. We can't use the "dev" reliably without the
>> reference count. We are opening up use-after-free.
> 
> My understanding is we don't grab a device from
> coresight_find_device_by_fwnode().  The callers only check whether the
> device is present on the bus; if it isn't, the driver defers probe.
> 
> This is similiar to coresight_find_csdev_by_fwnode(), which calls
> put_device(dev) to release refcnt immediately.  This is why I
> suggested the change, so the two functions behave consistently.
> 

I see, sorry. I saw some other uses of the device, but clearly I was 
wrong. May be we should simply re-structure the function to :

bool coresight_fwnode_device_present(fwnode)
{

	// find and drop the ref if required.
	return true/false;
}

The name "find_device_by_fwnode" and returning a freed reference doesn't
look good to me.

Suzuki

> Thanks,
> Leo

On Fri, Dec 19, 2025 at 11:48:35AM +0000, Suzuki K Poulose wrote:

[...]

> > My understanding is we don't grab a device from
> > coresight_find_device_by_fwnode().  The callers only check whether the
> > device is present on the bus; if it isn't, the driver defers probe.
> > 
> > This is similiar to coresight_find_csdev_by_fwnode(), which calls
> > put_device(dev) to release refcnt immediately.  This is why I
> > suggested the change, so the two functions behave consistently.
> > 
> 
> I see, sorry. I saw some other uses of the device, but clearly I was wrong.
> May be we should simply re-structure the function to :

No worries and thanks for confirmation.

> bool coresight_fwnode_device_present(fwnode)
> {
> 
> 	// find and drop the ref if required.
> 	return true/false;
> }
> 
> The name "find_device_by_fwnode" and returning a freed reference doesn't
> look good to me.

Renaming is good.  Maybe use a separate patch to rename:

  coresight_find_csdev_by_fwnode() -> coresight_fwnode_csdev_present()

Thanks,
Leo