Sorry to keep pestering about the patch series, but with the addition
of P_PIDFD, I react once again..

On Sat, Jul 27, 2019 at 1:53 AM Christian Brauner <[email protected]> wrote:
>
> --- a/kernel/exit.c
> +++ b/kernel/exit.c
> @@ -1555,6 +1555,7 @@ static long do_wait(struct wait_opts *wo)
>  static long kernel_waitid(int which, pid_t upid, struct waitid_info *infop,
>                           int options, struct rusage *ru)
>  {
> +       struct fd f;

Please don't do 'struct fd' at this level. That results in this ugly code later:

> -       put_pid(pid);
> +       if (which == P_PIDFD)
> +               fdput(f);
> +       else
> +               put_pid(pid);

which just looks nasty.

Instead, do all the 'file descriptor to pid' games here:

> +       case P_PIDFD:
> +               type = PIDTYPE_PID;
> +               if (upid < 0)
> +                       return -EINVAL;
> +
> +               f = fdget(upid);
> +               if (!f.file)
> +                       return -EBADF;
> +
> +               pid = pidfd_pid(f.file);
> +               if (IS_ERR(pid)) {
> +                       fdput(f);
> +                       return PTR_ERR(pid);
> +               }
>                 break;

and make thus just do something like

        pid = get_pid_from_fd(upid);
        if (IS_ERR(pid))
                return PTR_ERR(pid);

and now do that "fd to pid" in that helper function, and get the
reference to 'struct pid *' there instead.

Which you can actually do efficiently and lightly without even getting
a ref to the 'struct file'. Something like

  struct pid *fd_to_pid(unsigned int fd)
  {
        struct fd f;
        struct pid *pid;

        f = fdget(fd);
        if (!f.file)
                return ERR_PTR(-EBADF);
        pid = pidfd_pid(f.file);
        if (!IS_ERR(pid))
                get_pid(pid);
        fdput(f);
        return pid;
  }

is the stupid and straightforward thing, but if you want to be
*clever* you can actually avoid getting a reference to the 'struct
file *" entirely, and do the fd->pid lookup under rcu_read_lock()
instead. It's slightly more complex, but it avoids the fdget/fdput
reference count games entirely.

And then all that kernel_waitid() ever worries about is "struct pid
*", and the ending goes back to just that simple

        put_pid(pid);
        return ret;

instead.

This was kind of my point of doing all the "find_get_pid()" games in
the "switch()" statement - the different cases have different ways to
look up what the "struct pid *" pointer should be, but they should all
just look up a pid pointer, and then nothing else needs to care about
'type' any more. See?

Hmm?

                Linus

Reply via email to