Re: [PATCH 2/3] fd/locks: allow get the lock owner by F_OFD_GETLK

[Jeff Layton <jlayton@kernel.org>]

On Fri, 2023-06-23 at 22:18 +0500, stsp wrote:
> 23.06.2023 20:25, Christian Brauner пишет:
> > On Wed, Jun 21, 2023 at 07:05:12AM -0400, Jeff Layton wrote:
> > > On Wed, 2023-06-21 at 15:42 +0500, stsp wrote:
> > > > 21.06.2023 15:35, Jeff Layton пишет:
> > > > > I don't think we can change this at this point.
> > > > > 
> > > > > The bottom line (again) is that OFD locks are owned by the file
> > > > > descriptor (much like with flock()), and since file descriptors can be
> > > > > shared across multiple process it's impossible to say that some single
> > > > > process owns it.
> > > > What's the problem with 2 owners?
> > > > Can't you get one of them, rather than
> > > > meaningless -1?
> > > > Compare this situation with read locks.
> > > > They can overlap, so when you get an
> > > > info about a read lock (except for the
> > > > new F_UNLCK case), you get the info
> > > > about *some* of the locks in that range.
> > > > In the case of multiple owners, you
> > > > likewise get the info about about some
> > > > owner. If you iteratively send them a
> > > > "please release this lock" message
> > > > (eg in a form of SIGKILL), then you
> > > > traverse all, and end up with the
> > > > lock-free area.
> > > > Is there really any problem here?
> > > Yes. Ambiguous answers are worse than none at all.
> > I agree.
> > 
> > A few minor observations:
> > 
> > SCM_RIGHTS have already been mentioned multiple times. But I'm not sure
> > it's been mentioned explicitly but that trivially means it's possible to
> > send an fd to a completely separate thread-group, then kill off the
> > sending thread-group by killing their thread-group leader. Bad enough as
> > the identifier is now useless. But it also means that at some later
> > point that pid can be recycled.
> Come on.
> I never proposed anything like this.
> Of course the returned pid should be
> the pid of the current, actual owner,
> or one of current owners.
> If someone else proposed to return
> stalled pids, then it wasn't me.


Beyond all of this, there is a long history of problems with the l_pid
field as well with network filesystems, even with traditional POSIX
locks. What should go into the l_pid when a traditional POSIX lock is
held by a process on a separate host?

While POSIX mandates it, the l_pid is really sort of a "legacy" field
that is really just for informational purposes only nowadays. It might
have been a reliable bit of information back in the 1980's, but even
since the 90's it was suspect as a source of information.

Even if you _know_ you hold a traditional POSIX lock, be careful
trusting the information in that field.
-- 
Jeff Layton <jlayton@kernel.org>