Re: [PATCH v2 5/7] rpmsg: char: Introduce a rpmsg driver for the rpmsg char device

[Arnaud POULIQUEN <arnaud.pouliquen@foss.st.com>]

Hi,

On 4/22/21 6:36 PM, Mathieu Poirier wrote:
> On Thu, Apr 22, 2021 at 09:58:27AM +0200, Arnaud POULIQUEN wrote:
>> On 4/21/21 7:40 PM, Mathieu Poirier wrote:
>>> Good day Arnaud,
>>>
>>> On Tue, Apr 13, 2021 at 03:44:56PM +0200, Arnaud Pouliquen wrote:
>>>> A rpmsg char device allows to probe the endpoint device on a remote name
>>>> service announcement.
>>>>
>>>> With this patch the /dev/rpmsgX interface is created either by a user
>>>> application or by the remote firmware.
>>>>
>>>> Signed-off-by: Arnaud Pouliquen <arnaud.pouliquen@foss.st.com>
>>>>
>>>> ---
>>>> update from V1:
>>>>  - add missing unregister_rpmsg_driver call on module exit.
>>>>
>>>> ---
>>>>  drivers/rpmsg/rpmsg_char.c | 59 +++++++++++++++++++++++++++++++++++++-
>>>>  1 file changed, 58 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/rpmsg/rpmsg_char.c b/drivers/rpmsg/rpmsg_char.c
>>>> index a64249d83172..4606787b7011 100644
>>>> --- a/drivers/rpmsg/rpmsg_char.c
>>>> +++ b/drivers/rpmsg/rpmsg_char.c
>>>> @@ -26,6 +26,8 @@
>>>>  #include "rpmsg_char.h"
>>>>  #include "rpmsg_internal.h"
>>>>  
>>>> +#define RPMSG_CHAR_DEVNAME "rpmsg-raw"
>>>> +
>>>
>>> Why not simply call it rpmsg-char?
>>
>> I would avoid to link the rpmsg name service to the Linux Kernel device.
> 
> To me that's exactly what we want to do...  Am I missing something?
> 
>>
>>>
>>>>  static dev_t rpmsg_major;
>>>>  
>>>>  static DEFINE_IDA(rpmsg_ept_ida);
>>>> @@ -403,13 +405,67 @@ int rpmsg_chrdev_create_eptdev(struct rpmsg_device *rpdev, struct device *parent
>>>>  }
>>>>  EXPORT_SYMBOL(rpmsg_chrdev_create_eptdev);
>>>>  
>>>> +static int rpmsg_chrdev_probe(struct rpmsg_device *rpdev)
>>>> +{
>>>> +	struct rpmsg_channel_info chinfo;
>>>> +	struct rpmsg_eptdev *eptdev;
>>>> +
>>>> +	if (!rpdev->ept)
>>>> +		return -EINVAL;
>>>> +
>>>> +	memcpy(chinfo.name, RPMSG_CHAR_DEVNAME, sizeof(RPMSG_CHAR_DEVNAME));
>>>> +	chinfo.src = rpdev->src;
>>>> +	chinfo.dst = rpdev->dst;
>>>> +
>>>> +	eptdev = __rpmsg_chrdev_create_eptdev(rpdev, &rpdev->dev, chinfo, NULL);
>>>> +	if (IS_ERR(eptdev))
>>>> +		return PTR_ERR(eptdev);
>>>> +
>>>> +	/* Set the private field of the default endpoint to retrieve context on callback. */
>>>> +	rpdev->ept->priv = eptdev;
>>>
>>> This is already done in rpmsg_create_ept() when rpmsg_eptdev_open() is called.
>>>
>>>> +
>>>> +	return 0;
>>>> +}
>>>> +
>>>> +static void rpmsg_chrdev_remove(struct rpmsg_device *rpdev)
>>>> +{
>>>> +	int ret;
>>>> +
>>>> +	ret = device_for_each_child(&rpdev->dev, NULL, rpmsg_chrdev_destroy_eptdev);
>>>> +	if (ret)
>>>> +		dev_warn(&rpdev->dev, "failed to destroy endpoints: %d\n", ret);
>>>> +}
>>>> +
>>>> +static struct rpmsg_device_id rpmsg_chrdev_id_table[] = {
>>>> +	{ .name	= RPMSG_CHAR_DEVNAME },
>>>> +	{ },
>>>> +};
>>>> +
>>>> +static struct rpmsg_driver rpmsg_chrdev_driver = {
>>>> +	.probe = rpmsg_chrdev_probe,
>>>> +	.remove = rpmsg_chrdev_remove,
>>>> +	.id_table = rpmsg_chrdev_id_table,
>>>> +	.callback = rpmsg_ept_cb,
>>>
>>> Not sure why we need a callback associated to this driver when
>>> rpmsg_eptdev_open() already creates and rpmsg_endpoint.  To me the only thing
>>> having a callback provides is the association between the rpmsg_device and the
>>> rpmsg_endpoint[1] that happens in rpmsg_dev_probe().  The QC folks already do
>>> this association in their platform code[2].  Since this is not done in
>>> __rpmsg_create_ept() a check for rpdev->ept == NULL could be done in
>>> rpmsg_eptdev_open() and do the assignment there. 
>>>
>>> [1]. https://elixir.bootlin.com/linux/v5.12-rc6/source/drivers/rpmsg/rpmsg_core.c#L513  
>>> [2]. https://elixir.bootlin.com/linux/v5.12-rc6/source/drivers/rpmsg/qcom_glink_native.c#L1623
>>>
>>
>> That's a good point! When I started the redesign, I faced some issues with the
>> approach you propose. But as I can not remember the reason and because the code
>> has evolved, i need to re-think about this.
>>
> 
> Glad to see we're on the same page.  I stared at this code for a very long time,
> thinking there was some kind of bigger picture I wasn't getting.

I finally found the time to investigate this. If I remember now why I used this
approach, I also saw that my patchset does not work with the QCOM platform driver.

As a first step of explanation, let's ignore the QC platform.
rpdev->ept is null for the rpmsg ctrldev device created by the virtio rpmsg bus.
If no default endpoint is created on rpmsg_chrdev_probe, it is not possible to
differentiate the two in rpmsg_eptdev_open based on rpdev->ept == NULL.

Now let's add the QC implementation
As you mentioned, QC sets the rpdev->ept to a default endpoint before
registering the rpmsg ctrldev. This shows that it is not reasonable to expect to
handle all use cases based on the rpdev->ept value.

So, to summarize, I need to rework this, probably by adding a new field in the
rpmsg_eptdev structure, to properly handle the endpoint creation in the
rpmsg_eptdev_open function.

Regards,
Arnaud

> 
> 
>> Thanks,
>> Arnaud
>>
>>
>>>> +	.drv = {
>>>> +		.name = "rpmsg_chrdev",
>>>> +	},
>>>> +};
>>>> +
>>>>  static int rpmsg_chrdev_init(void)
>>>>  {
>>>>  	int ret;
>>>>  
>>>>  	ret = alloc_chrdev_region(&rpmsg_major, 0, RPMSG_DEV_MAX, "rpmsg_char");
>>>> -	if (ret < 0)
>>>> +	if (ret < 0) {
>>>>  		pr_err("rpmsg: failed to allocate char dev region\n");
>>>> +		return ret;
>>>> +	}
>>>> +
>>>> +	ret = register_rpmsg_driver(&rpmsg_chrdev_driver);
>>>> +	if (ret < 0) {
>>>> +		pr_err("rpmsg: failed to register rpmsg raw driver\n");
>>>> +		unregister_chrdev_region(rpmsg_major, RPMSG_DEV_MAX);
>>>> +	}
>>>>  
>>>>  	return ret;
>>>>  }
>>>> @@ -417,6 +473,7 @@ postcore_initcall(rpmsg_chrdev_init);
>>>>  
>>>>  static void rpmsg_chrdev_exit(void)
>>>>  {
>>>> +	unregister_rpmsg_driver(&rpmsg_chrdev_driver);
>>>>  	unregister_chrdev_region(rpmsg_major, RPMSG_DEV_MAX);
>>>>  }
>>>>  module_exit(rpmsg_chrdev_exit);
>>>> -- 
>>>> 2.17.1
>>>>