Re: [PATCH net-next v2 2/3] ptp: support multiple timestamp event readers

[Vinicius Costa Gomes <vinicius.gomes@intel.com>]

Xabier Marquiegui <xabier.marquiegui@gmail.com> writes:

>> Using the pid of the task will break when using some form of file
>> descriptor passing. e.g. Task A opened the chardev, called the ioctl()
>> with some mask and then passed the fd to Task B, that's actually going
>> to use the fd.
>> 
>> Is this something that we want to support? Am I missing something?
>
> Thank you very much for your valuable comments Vinicius,
>
> Let me try to confirm if I understand what you are observing here.
>
> Let's say we have a process with two threads. One of them opens the device
> file to do IOCTL operations, and the other one opens the device file to
> read timestamp events. Since both have the same PID, all the operations
> (read,release...) would fail because I designed them under the assumption
> that only one open operation would happen per PID.
>

I was initially thinking about another scenario: when Process A passes
the open file descriptor via SCM_RIGHTS to Process B.

But your scenario also shows the limitations of the current approach.

> This is indeed not as safe as it should be and I should try to improve it.
>
> I am already looking at alternatives, but maybe someone can give me a hint.
> Do you have any suggestions on what I could do to have file operations 
> (read, release...) determine which open instance they belong to?

Taking a quick look, it seems that you would have to change 'struct
posix_clock_file_operations' to also pass around the 'struct file' of
the file being used.

That way we can track each user/"open()". And if one program decides
that it needs to have have multiple fds with different masks, and so
different queues, it should just work.

What do you think?


Cheers,
-- 
Vinicius