Re: [PATCH v1 3/4] signal: support CLONE_PIDFD with pidfd_send_signal

From: Oleg Nesterov
Date: Wed Apr 17 2019 - 11:21:16 EST


On 04/17, Christian Brauner wrote:
>
> On Wed, Apr 17, 2019 at 04:01:06PM +0200, Oleg Nesterov wrote:
> > On 04/16, Christian Brauner wrote:
> > >
> > > @@ -3581,12 +3588,12 @@ SYSCALL_DEFINE4(pidfd_send_signal, int, pidfd, int, sig,
> > > if (flags)
> > > return -EINVAL;
> > >
> > > - f = fdget_raw(pidfd);
> > > + f = fdget(pidfd);
> >
> > could you explain this change?
> >
> > I am just curious, I don't understand why should we disallow O_PATH and how
> > this connects to this patch.
>
> Sending a signal through a pidfd is considered to be on a par with a
> "write" to that pidfd.

OK, but how this connects to "support pidfds" ?

> Additionally, we use the fops associated with the fd to detect whether
> it is actually a pidfd or not. This is not possible with O_PATH since
> f_ops will be set to dummy fops.

indeed... I didn't know this, thanks!

But this means that pidfd_send_signal() will return -EBADF with or without
this change; pidfd_to_pid() will return -EBADF even if fdget_raw() suceeds,
right?

To clarify, I am not arguing. I am trying to understand why exactly do we
need this s/fdget_raw/fdget/ change and, why it doesn't come as a separate
patch. Can you add a note into the changelog?

Oleg.