diff mbox series

[v3,06/13] coresight: Fix loss of connection info when a module is unloaded

Message ID 20230329115329.2747724-7-james.clark@arm.com (mailing list archive)
State New, archived
Headers show
Series coresight: Fix CTI module refcount leak by making it a helper device | expand

Commit Message

James Clark March 29, 2023, 11:53 a.m. UTC
child_fwnode should be a read only property based on the DT. If it's
cleared on the parent device when a child is unloaded, then when the
child is loaded again the connection won't be remade.

child_dev should be cleared instead which signifies that the connection
should be remade when the child_fwnode registers a new coresight_device.

Similarly the reference count shouldn't be decremented as long as the
parent device exists. The correct place to drop the reference is in
coresight_release_platform_data() which is already done.

Signed-off-by: James Clark <james.clark@arm.com>
---
 drivers/hwtracing/coresight/coresight-core.c | 10 ++--------
 1 file changed, 2 insertions(+), 8 deletions(-)

Comments

Suzuki K Poulose March 30, 2023, 12:42 p.m. UTC | #1
Hi James

On 29/03/2023 12:53, James Clark wrote:
> child_fwnode should be a read only property based on the DT. If it's
> cleared on the parent device when a child is unloaded, then when the
> child is loaded again the connection won't be remade.
> 
> child_dev should be cleared instead which signifies that the connection
> should be remade when the child_fwnode registers a new coresight_device.
> 
> Similarly the reference count shouldn't be decremented as long as the
> parent device exists. The correct place to drop the reference is in
> coresight_release_platform_data() which is already done.
> 
> Signed-off-by: James Clark <james.clark@arm.com>

This looks like a bug in the existing driver. Please could you keep this
at the beginning of the series, so that it is easier to backport for
stable ?

Also, please add :

Fixes: 37ea1ffddffa ("coresight: Use fwnode handle instead of device names")
Fixes: 2af89ebacf29 ("coresight: Clear the connection field properly")


> ---
>   drivers/hwtracing/coresight/coresight-core.c | 10 ++--------
>   1 file changed, 2 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/hwtracing/coresight/coresight-core.c b/drivers/hwtracing/coresight/coresight-core.c
> index bfe1b93aca4e..12cbb68e8e1c 100644
> --- a/drivers/hwtracing/coresight/coresight-core.c
> +++ b/drivers/hwtracing/coresight/coresight-core.c
> @@ -1414,13 +1414,8 @@ static int coresight_remove_match(struct device *dev, void *data)
>   		if (csdev->dev.fwnode == conn->dest_fwnode) {
>   			iterator->orphan = true;
>   			coresight_remove_links(iterator, conn);
> -			/*
> -			 * Drop the reference to the handle for the remote
> -			 * device acquired in parsing the connections from
> -			 * platform data.
> -			 */
> -			fwnode_handle_put(conn->dest_fwnode);
> -			conn->dest_fwnode = NULL;
> +
> +			conn->dest_dev = NULL;
>   			/* No need to continue */
>   			break;
>   		}
> @@ -1553,7 +1548,6 @@ void coresight_release_platform_data(struct coresight_device *csdev,
>   		 * is going away
>   		 */
>   		fwnode_handle_put(conns[i].dest_fwnode);
> -		pdata->out_conns[i].dest_fwnode = NULL;

This change is not required, as we are freeing the pdata anyway.
Keeping this might be beneficial if someone else comes looking
for references while we are going out. I don't see how we could
hit that, but just to be safe.

Suzuki
Suzuki K Poulose March 30, 2023, 2:01 p.m. UTC | #2
On 30/03/2023 13:42, Suzuki K Poulose wrote:
> Hi James
> 
> On 29/03/2023 12:53, James Clark wrote:
>> child_fwnode should be a read only property based on the DT. If it's
>> cleared on the parent device when a child is unloaded, then when the
>> child is loaded again the connection won't be remade.
>>
>> child_dev should be cleared instead which signifies that the connection
>> should be remade when the child_fwnode registers a new coresight_device.
>>
>> Similarly the reference count shouldn't be decremented as long as the
>> parent device exists. The correct place to drop the reference is in
>> coresight_release_platform_data() which is already done.
>>
>> Signed-off-by: James Clark <james.clark@arm.com>
> 
> This looks like a bug in the existing driver. Please could you keep this
> at the beginning of the series, so that it is easier to backport for
> stable ?
> 
> Also, please add :
> 
> Fixes: 37ea1ffddffa ("coresight: Use fwnode handle instead of device 
> names")
> Fixes: 2af89ebacf29 ("coresight: Clear the connection field properly")
> 
> 
>> ---
>>   drivers/hwtracing/coresight/coresight-core.c | 10 ++--------
>>   1 file changed, 2 insertions(+), 8 deletions(-)
>>
>> diff --git a/drivers/hwtracing/coresight/coresight-core.c 
>> b/drivers/hwtracing/coresight/coresight-core.c
>> index bfe1b93aca4e..12cbb68e8e1c 100644
>> --- a/drivers/hwtracing/coresight/coresight-core.c
>> +++ b/drivers/hwtracing/coresight/coresight-core.c
>> @@ -1414,13 +1414,8 @@ static int coresight_remove_match(struct device 
>> *dev, void *data)
>>           if (csdev->dev.fwnode == conn->dest_fwnode) {
>>               iterator->orphan = true;
>>               coresight_remove_links(iterator, conn);
>> -            /*
>> -             * Drop the reference to the handle for the remote
>> -             * device acquired in parsing the connections from
>> -             * platform data.
>> -             */
>> -            fwnode_handle_put(conn->dest_fwnode);
>> -            conn->dest_fwnode = NULL;
>> +
>> +            conn->dest_dev = NULL;
>>               /* No need to continue */
>>               break;
>>           }
>> @@ -1553,7 +1548,6 @@ void coresight_release_platform_data(struct 
>> coresight_device *csdev,
>>            * is going away
>>            */
>>           fwnode_handle_put(conns[i].dest_fwnode);
>> -        pdata->out_conns[i].dest_fwnode = NULL;
> 
> This change is not required, as we are freeing the pdata anyway.
> Keeping this might be beneficial if someone else comes looking
> for references while we are going out. I don't see how we could
> hit that, but just to be safe.

For reference, I could reproduce the problem by :

On juno, load all coresight modules.

$ cd /sys/bus/coresight/devices/
$ echo 1 > tmc_etr0/enable_sink
$ echo 1 > etm0/enable_source

Works fine ^

$ echo 0 > etm0/enable_source
$ rmmod coresight-funnel
$ modprobe coresight-funnel
$ echo 1 > etm0/enable_source
-bash: echo: write error: Invalid argument

Suzuki

> 
> Suzuki
>
diff mbox series

Patch

diff --git a/drivers/hwtracing/coresight/coresight-core.c b/drivers/hwtracing/coresight/coresight-core.c
index bfe1b93aca4e..12cbb68e8e1c 100644
--- a/drivers/hwtracing/coresight/coresight-core.c
+++ b/drivers/hwtracing/coresight/coresight-core.c
@@ -1414,13 +1414,8 @@  static int coresight_remove_match(struct device *dev, void *data)
 		if (csdev->dev.fwnode == conn->dest_fwnode) {
 			iterator->orphan = true;
 			coresight_remove_links(iterator, conn);
-			/*
-			 * Drop the reference to the handle for the remote
-			 * device acquired in parsing the connections from
-			 * platform data.
-			 */
-			fwnode_handle_put(conn->dest_fwnode);
-			conn->dest_fwnode = NULL;
+
+			conn->dest_dev = NULL;
 			/* No need to continue */
 			break;
 		}
@@ -1553,7 +1548,6 @@  void coresight_release_platform_data(struct coresight_device *csdev,
 		 * is going away
 		 */
 		fwnode_handle_put(conns[i].dest_fwnode);
-		pdata->out_conns[i].dest_fwnode = NULL;
 	}
 	if (csdev)
 		coresight_remove_conns_sysfs_group(csdev);