Re: [PATCH v2] signal: add procfd_signal() syscall

[Florian Weimer <fweimer@redhat.com>]

* Christian Brauner:

> On Mon, Dec 03, 2018 at 05:57:51PM +0100, Florian Weimer wrote:
>> * Christian Brauner:
>> 
>> > Ok, I finally have access to source code again. Scratch what I said above!
>> > I looked at the code and tested it. If the process has exited but not
>> > yet waited upon aka is a zombie procfd_send_signal() will return 0. This
>> > is identical to kill(2) behavior. It should've been sort-of obvious
>> > since when a process is in zombie state /proc/<pid> will still be around
>> > which means that struct pid must still be around.
>> 
>> Should we make this state more accessible, by providing a different
>> error code?
>
> No, I don't think we want that. Imho, It's not really helpful. Signals
> are still delivered to zombies. If zombie state were to always mean that
> no-one is going to wait on this thread anymore then it would make sense
> to me. But given that zombie can also mean that someone put a
> sleep(1000) right before their wait() call in the parent it seems odd to
> report back that it is a zombie.

It allows for error checking that the recipient of a signal is still
running.  It's obviously not reliable, but I think it could be helpful
in the context of closely cooperating processes.

>> Will the system call ever return ESRCH, given that you have a handle for
>> the process?
>
> Yes, whenever you signal a process that has already been waited upon:
> - get procfd handle referring to <proc>
> - <proc> exits and is waited upon
> - procfd_send_signal(procfd, ...) returns -1 with errno == ESRCH

I see, thanks.

>> Do you want to land all this in one kernel release?  I wonder how
>> applications are supposed to discover kernel support if functionality is
>> split across several kernel releases.  If you get EINVAL or EBADF, it
>> may not be obvious what is going on.
>
> Sigh, I get that but I really don't want to have to land this in one big
> chunk. I want this syscall to go in in a as soon as we can to fulfill
> the most basic need: having a way that guarantees us that we signal the
> process that we intended to signal.
>
> The thread case is easy to implement on top of it. But I suspect we will
> quibble about the exact semantics for a long time. Even now we have been
> on multiple - justified - detrous. That's all pefectly fine and
> expected. But if we have the basic functionality in we have time to do
> all of that. We might even land it in the same kernel release still. I
> really don't want to come of as tea-party-kernel-conservative here but I
> have time-and-time again seen that making something fancy and cover ever
> interesting feature in one patchset takes a very very long time.
>
> If you care about userspace being able to detect that case I can return
> EOPNOTSUPP when a tid descriptor is passed.

I suppose that's fine.  Or alternatively, when thread group support is
added, introduce a flag that applications have to use to enable it, so
that they can probe for support by checking support for the flag.

I wouldn't be opposed to a new system call like this either:

  int procfd_open (pid_t thread_group, pid_t thread_id, unsigned flags);

But I think this is frowned upon on the kernel side.

>> What happens if you use the new interface with an O_PATH descriptor?
>
> You get EINVAL. When an O_PATH file descriptor is created the kernel
> will set file->f_op = &empty_fops at which point the check I added 
>         if (!proc_is_tgid_procfd(f.file))
>                 goto err;
> will fail. Imho this is correct behavior since technically signaling a
> struct pid is the equivalent of writing to a file and hence doesn't
> purely operate on the file descriptor level.

Yes, that's quite reasonable.  Thanks.

Florian