Re: [PATCH net v2 1/2] ptp: ocp: adjust sysfs entries to expose tty information

[Vadim Fedorenko <vadim.fedorenko@linux.dev>]

On 14/08/2024 11:30, Jiri Slaby wrote:
> On 14. 08. 24, 10:37, Vadim Fedorenko wrote:
>> On 14/08/2024 06:00, Jiri Slaby wrote:
>>> On 13. 08. 24, 20:24, Vadim Fedorenko wrote:
>>>> On 13/08/2024 10:33, Greg Kroah-Hartman wrote:
>>>>> On Mon, Aug 05, 2024 at 03:04:59PM -0700, Vadim Fedorenko wrote:
>>>>>> Starting v6.8 the serial port subsystem changed the hierarchy of 
>>>>>> devices
>>>>>> and symlinks are not working anymore. Previous discussion made it 
>>>>>> clear
>>>>>> that the idea of symlinks for tty devices was wrong by design. 
>>>>>> Implement
>>>>>> additional attributes to expose the information. Fixes tag points 
>>>>>> to the
>>>>>> commit which introduced the change.
>>>>>>
>>>>>> Fixes: b286f4e87e32 ("serial: core: Move tty and serdev to be 
>>>>>> children of serial core port device")
>>>>>> Signed-off-by: Vadim Fedorenko <vadfed@meta.com>
>>>>>> ---
>>>>>>   drivers/ptp/ptp_ocp.c | 68 ++++++++++++++++++++++++++++++++ 
>>>>>> +----------
>>>>>>   1 file changed, 52 insertions(+), 16 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/ptp/ptp_ocp.c b/drivers/ptp/ptp_ocp.c
>>>>>> index ee2ced88ab34..7a5026656452 100644
>>>>>> --- a/drivers/ptp/ptp_ocp.c
>>>>>> +++ b/drivers/ptp/ptp_ocp.c
>>>>>> @@ -3346,6 +3346,55 @@ static EXT_ATTR_RO(freq, frequency, 1);
>>>>>>   static EXT_ATTR_RO(freq, frequency, 2);
>>>>>>   static EXT_ATTR_RO(freq, frequency, 3);
>>>>>> +static ssize_t
>>>>>> +ptp_ocp_tty_show(struct device *dev, struct device_attribute 
>>>>>> *attr, char *buf)
>>>>>> +{
>>>>>> +    struct dev_ext_attribute *ea = to_ext_attr(attr);
>>>>>> +    struct ptp_ocp *bp = dev_get_drvdata(dev);
>>>>>> +    struct ptp_ocp_serial_port *port;
>>>>>> +
>>>>>> +    port = (void *)((uintptr_t)bp + (uintptr_t)ea->var);
>>>>>
>>>>> That's insane pointer math, how do we know this is correct?
>>>>>
>>>>>> +    return sysfs_emit(buf, "ttyS%d", port->line);
>>>>>> +}
>>>>>> +
>>>>>> +static umode_t
>>>>>> +ptp_ocp_timecard_tty_is_visible(struct kobject *kobj, struct 
>>>>>> attribute *attr, int n)
>>>>>> +{
>>>>>> +    struct ptp_ocp *bp = dev_get_drvdata(kobj_to_dev(kobj));
>>>>>> +    struct ptp_ocp_serial_port *port;
>>>>>> +    struct device_attribute *dattr;
>>>>>> +    struct dev_ext_attribute *ea;
>>>>>> +
>>>>>> +    if (strncmp(attr->name, "tty", 3))
>>>>>> +        return attr->mode;
>>>>>> +
>>>>>> +    dattr = container_of(attr, struct device_attribute, attr);
>>>>>> +    ea = container_of(dattr, struct dev_ext_attribute, attr);
>>>>>> +    port = (void *)((uintptr_t)bp + (uintptr_t)ea->var);
>>>>>
>>>>> That's crazy pointer math, how are you ensured that it is correct?  
>>>>> Why
>>>>> isn't there a container_of() thing here instead?
>>>>
>>>> Well, container_of cannot be used here because the attributes are 
>>>> static
>>>> while the function reads dynamic instance. The only values that are
>>>> populated into the attributes of the group are offsets.
>>>> But I can convert it to a helper which will check that the offset 
>>>> provided is the real offset of the structure we expect. And it could 
>>>> be reused in both "is_visible" and "show" functions.
>>>
>>> Strong NACK against this approach.
>>>
>>> What about converting those 4 ports into an array and adding an enum 
>>> { PORT_GNSS, POTR_GNSS2, PORT_MAC, PORT_NMEA }?
>>
>> Why is it a problem? I don't see big difference between these 2
>> implementations:
>>
>> struct ptp_ocp_serial_port *get_port(struct ptp_ocp *bp, void *offset)
>> {
>>      switch((uintptr_t)offset) {
>>          case offsetof(struct ptp_ocp, gnss_port):
>>              return &bp->gnss_port;
>>          case offsetof(struct ptp_ocp, gnss2_port):
>>              return &bp->gnss2_port;
>>          case offsetof(struct ptp_ocp, mac_port):
>>              return &bp->mac_port;
>>          case offsetof(struct ptp_ocp, nmea_port):
>>              return &bp->nmea_port;
>>      }
>>      return NULL;
>> }
>>
>> and:
>>
>> struct ptp_ocp_serial_port *get_port(struct ptp_ocp *bp, void *offset)
>> {
>>      switch((enum port_type)offset) {
>>          case PORT_GNSS:
>>              return &bp->tty_port[PORT_GNSS];
>>          case PORT_GNSS2:
>>              return &bp->tty_port[PORT_GNSS2];
>>          case PORT_MAC:
>>              return &bp->tty_port[PORT_MAC];
>>          case PORT_NMEA:
>>              return &bp->tty_port[PORT_NMEA];
>>      }
>>      return NULL;
>> }
>>
>> The second option will require more LoC to change the initialization
>> part of the driver, but will not simplify the access.
>> If you suggest to use enum value directly, without the check, then
>> it will not solve the problem of checking the boundary, which Greg
>> refers to AFAIU.
> 
> Why do you need this get_port() here at all? Simply doing bp- 
>  >tty_port[ea->var] as in already present attrs will do the job, right?
> 

Alright, the v3 is up for review with this way applied,
thanks!

> thanks,