Re: [PATCH 2/4] mm: introduce external memory hinting API

From: Daniel Colascione <dancol@google.com>
To: Christian Brauner <christian.brauner@ubuntu.com>
Cc: Minchan Kim <minchan@kernel.org>,
	Kirill Tkhai <ktkhai@virtuozzo.com>,
	 Michal Hocko <mhocko@suse.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	 LKML <linux-kernel@vger.kernel.org>,
	linux-mm <linux-mm@kvack.org>,
	 Linux API <linux-api@vger.kernel.org>,
	oleksandr@redhat.com,  Suren Baghdasaryan <surenb@google.com>,
	Tim Murray <timmurray@google.com>,
	 Sandeep Patil <sspatil@google.com>,
	Sonny Rao <sonnyrao@google.com>,
	 Brian Geffon <bgeffon@google.com>,
	Johannes Weiner <hannes@cmpxchg.org>,
	 Shakeel Butt <shakeelb@google.com>,
	John Dias <joaodias@google.com>
Subject: Re: [PATCH 2/4] mm: introduce external memory hinting API
Date: Mon, 13 Jan 2020 13:04:44 -0800	[thread overview]
Message-ID: <CAKOZueu=U4c2URaq8Pz-B00XV+TxaKwHRNXv3BUiDbQrLQpJ3A@mail.gmail.com> (raw)
In-Reply-To: <20200113204237.ew6nn4ohxu7auw3u@wittgenstein>

On Mon, Jan 13, 2020 at 12:42 PM Christian Brauner
<christian.brauner@ubuntu.com> wrote:
>
> On Mon, Jan 13, 2020 at 11:27:03AM -0800, Daniel Colascione wrote:
> > On Mon, Jan 13, 2020 at 11:10 AM Christian Brauner
> > <christian.brauner@ubuntu.com> wrote:
> > > This does not
> > > affect the permission checking you're performing here.
> >
> > Pidfds-as-capabilities sounds like a good change. Can you clarify what
> > you mean here though? Do you mean that in order to perform some
> > process-directed operation X on process Y, the pidfd passed to X must
> > have been opened with PIDFD_CAP_X *and* the process *using* the pidfds
> > must be able to perform operation X on process Y? Or do pidfds in this
> > model "carry" permissions in the same way that an ordinary file
> > descriptor "carries" the ability to write to a file if it was opened
> > with O_WRONLY even if the FD is passed to a process that couldn't
> > otherwise write to that file? Right now, pidfds are identity-only and
> > always rely on the caller's permissions. I like the capability bit
> > model because it makes pidfds more consistent with other file
> > descriptors and enabled delegation of capabilities across the system.
>
> I'm going back and forth on this. My initial implementation has it that
> you'd need both, PIDFD_FLAG/CAP_X and the process using the pidfd must
> be able to perform the operation X on process Y. The alternative becomes
> tricky for e.g. anything that requires ptrace_may_access() permissions
> such as getting an fd out from another task based on its pidfd and so
> on.

I think the alternative is necessary though. What's the point of the
pidfd capability bits if they don't grant access? If I have a pidfd
for Y that doesn't let me do operation X, but I have ambient authority
to do Y anyway, then I can just make my own pidfd for Y and then use
that new pidfd to do X. AFAICT, pidfd capabilities only do something
when they replace ptrace_may_access and friends for access control.
Otherwise, they seem purely advisory. Am I missing something?