Re: [PATCH 3/4] rpmsg: ctrl: Add check on rpmsg device removability from user space

[Arnaud POULIQUEN]

On 6/17/21 10:02 AM, Arnaud POULIQUEN wrote:
> Hello Mathieu,
> 
> On 6/16/21 7:15 PM, Mathieu Poirier wrote:
>> On Wed, Jun 16, 2021 at 11:30:51AM +0200, Arnaud POULIQUEN wrote:
>>>
>>>
>>> On 6/15/21 7:46 PM, Mathieu Poirier wrote:
>>>> On Fri, Jun 04, 2021 at 11:14:05AM +0200, Arnaud Pouliquen wrote:
>>>>> Using the RPMSG_RELEASE_DEV_IOCTL is possible to remove any
>>>>> rpmsg device (such as the rpmsg ns or the rpmsg ctrldev).
>>>>>
>>>>> Add a new field to store the removability of the device.
>>>>>
>>>>> By default the rpmsg device can not be removed by user space. It is
>>>>> set to 1 by the rpmsg ctrl on RPMSG_CREATE_DEV_IOCTL request, but
>>>>> could also be set by an rpmsg driver during probe.
>>>>>
>>>>> Signed-off-by: Arnaud Pouliquen <arnaud.pouliquen@xxxxxxxxxxx>
>>>>> ---
>>>>>  drivers/rpmsg/rpmsg_ctrl.c | 17 ++++++++++++++++-
>>>>>  include/linux/rpmsg.h      |  2 ++
>>>>>  2 files changed, 18 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/drivers/rpmsg/rpmsg_ctrl.c b/drivers/rpmsg/rpmsg_ctrl.c
>>>>> index cb19e32d05e1..e93c6ec49038 100644
>>>>> --- a/drivers/rpmsg/rpmsg_ctrl.c
>>>>> +++ b/drivers/rpmsg/rpmsg_ctrl.c
>>>>> @@ -74,6 +74,7 @@ static long rpmsg_ctrldev_ioctl(struct file *fp, unsigned int cmd,
>>>>>  	struct rpmsg_endpoint_info eptinfo;
>>>>>  	struct rpmsg_channel_info chinfo;
>>>>>  	struct rpmsg_device *rpdev;
>>>>> +	struct device *dev;
>>>>>  	int ret = 0;
>>>>>  
>>>>>  	if (copy_from_user(&eptinfo, argp, sizeof(eptinfo)))
>>>>> @@ -95,11 +96,25 @@ static long rpmsg_ctrldev_ioctl(struct file *fp, unsigned int cmd,
>>>>>  		if (!rpdev) {
>>>>>  			dev_err(&ctrldev->dev, "failed to create %s channel\n", chinfo.name);
>>>>>  			ret = -ENXIO;
>>>>> +		} else {
>>>>> +			/* Allow user space to release the device. */
>>>>> +			rpdev->us_removable = 1;
>>>>
>>>> As a rule of thumb I try really hard to avoid introducing new flags.  In this case we
>>>> can attain the same result by looking at chinfo->name, chinfo->src and
>>>> chinfo->dst.  I would introduce a new inline function in rpmsg_internal.h,
>>>> something like rpmsg_chrdev_is_ctrl_dev(), and compare the specifics in chinfo
>>>> to rpdev->id.name, rpdev->src and rpdev->dst.  If they all match then the
>>>> operation is refused.
>>>
>>> Something must have escaped me, because i turn around your your proposal,
>>> without understand it.
>>>
>>> The "us_removable" flag is not only for the rpmsg_ctrl, but for any rpmsg device
>>> that have not to be released by user application. Either because there are core
>>> ( rpmsg_ctrl, rpmsg_ns) or because a rpmsg driver don't allow to unbind its
>>> rpmsg devices.
>>>
>>
>> I don't see how the current patch would allow a driver to prevent user space
>> from releasing a rpmsg device since the sysfs attribute can be changed at will.
>> So even if the driver sets the flag user space can still revert it.
> 
> 
> The patch [4/4] define the a read only attribute using the rpmsg_show_attr
> declaration[1]. So the userspace can't change it.
> 
> This also has the advantage of not allowing the new IOCTRL API to be used by
> default for legacy RPMSg devices without a specific patch.
> 
> [1] https://elixir.bootlin.com/linux/latest/source/drivers/rpmsg/rpmsg_core.c#L362
> 
>>
>>> look to me that rpmsg_chrdev_is_ctrl_dev just prevents rpmsg ctrl to be released
>>> by the RPMSG_RELEASE_DEV_IOCTL.
>>
>> That is correct.  I did not address rpmsg_ns to keep things simple but it would
>> also have to be handled properly.
>>
>>>
>>> Please, could you clarify what you have in mind here?
>>
>> Other than rpmsg_ctrl and rpmsg_ns I don't think we should introduce any
>> mechanism to prevent users from releasing an rpmsg.  Doing so needs root access
>> - if a user space process with root privileges can't be trusted then we have
>> bigger problems than unwanted releases of registered rpmsg devices.
> 
> That's make sense. If we go on this way we could also trust the root application
> for the rpmsg_ns and only protect the rpmsg_ctrl which can not release itself,
> as you proposed.

As discussed in the OpenAMP by-weekly meeting, I will send a new revision,
without the attribute.

Thanks,
Arnaud

> 
> Thanks,
> 
> Arnaud
> 
>>
>>>
>>> Thanks,
>>> Arnaud
>>>
>>>>
>>>> That way we don't introduce a new flag and there is also no need to call
>>>> rpmsg_find_device() twice.
>>>
>>>
>>>
>>>>
>>>> Thanks,
>>>> Mathieu
>>>>
>>>>>  		}
>>>>>  		break;
>>>>>  
>>>>>  	case RPMSG_RELEASE_DEV_IOCTL:
>>>>> -		ret = rpmsg_release_channel(ctrldev->rpdev, &chinfo);
>>>>> +		dev = rpmsg_find_device(ctrldev->rpdev->dev.parent, &chinfo);
>>>>> +		if (!dev)
>>>>> +			ret =  -ENXIO;
>>>>> +
>>>>> +		/* Verify that rpmsg device removal is allowed. */
>>>>> +		if (!ret) {
>>>>> +			rpdev = to_rpmsg_device(dev);
>>>>> +			if (!rpdev->us_removable)
>>>>> +				ret = -EACCES;
>>>>> +		}
>>>>> +		if (!ret)
>>>>> +			ret = rpmsg_release_channel(ctrldev->rpdev, &chinfo);
>>>>>  		if (ret)
>>>>>  			dev_err(&ctrldev->dev, "failed to release %s channel (%d)\n",
>>>>>  				chinfo.name, ret);
>>>>> diff --git a/include/linux/rpmsg.h b/include/linux/rpmsg.h
>>>>> index d97dcd049f18..3642aad1a789 100644
>>>>> --- a/include/linux/rpmsg.h
>>>>> +++ b/include/linux/rpmsg.h
>>>>> @@ -47,6 +47,7 @@ struct rpmsg_channel_info {
>>>>>   * @ept: the rpmsg endpoint of this channel
>>>>>   * @announce: if set, rpmsg will announce the creation/removal of this channel
>>>>>   * @little_endian: True if transport is using little endian byte representation
>>>>> + * @us_removable: True if userspace application has permission to remove the rpmsg device
>>>>>   */
>>>>>  struct rpmsg_device {
>>>>>  	struct device dev;
>>>>> @@ -57,6 +58,7 @@ struct rpmsg_device {
>>>>>  	struct rpmsg_endpoint *ept;
>>>>>  	bool announce;
>>>>>  	bool little_endian;
>>>>> +	bool us_removable;
>>>>>  
>>>>>  	const struct rpmsg_device_ops *ops;
>>>>>  };
>>>>> -- 
>>>>> 2.17.1
>>>>>