Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20220202151734.10418-1-jandryuk@gmail.com> <YfrC/fQvQ1nkIJNJ@grain>
In-Reply-To: <YfrC/fQvQ1nkIJNJ@grain>
From:   Jason Andryuk <jandryuk@gmail.com>
Date:   Wed, 2 Feb 2022 14:48:48 -0500
Message-ID: <CAKf6xpsfDpEpjSjYbGU5fQOxqjccbu3CAXTxpcKggKYu6qEfNA@mail.gmail.com>
Subject: Re: [PATCH] kcmp: Comment get_file_raw_ptr() RCU usage
To:     Cyrill Gorcunov <gorcunov@gmail.com>
Cc:     open list <linux-kernel@vger.kernel.org>,
        "Eric W . Biederman" <ebiederm@xmission.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Feb 2, 2022 at 12:44 PM Cyrill Gorcunov <gorcunov@gmail.com> wrote:
>
> On Wed, Feb 02, 2022 at 10:17:34AM -0500, Jason Andryuk wrote:
> > This usage of RCU appears wrong since the pointer is passed outside the
> > RCU region.  However, it is not dereferenced, so it is "okay".  Leave a
> > comment for the next reader.
> >
> > Without a reference, these comparisons are racy, but even with their use
> > inside an RCU region, the result could go stale.
> >
> > Signed-off-by: Jason Andryuk <jandryuk@gmail.com>
> > ---
> > I was looking for examples of task_lookup_fd_rcu()/files_lookup_fd_rcu()
> > and found this.  It differed from the example given in
> > Documentation/filesystems/files.rst, so I was initially confused.  A
> > comment seemed appropriate to avoid confusion.
> >
> >  kernel/kcmp.c | 3 +++
> >  1 file changed, 3 insertions(+)
> >
> > diff --git a/kernel/kcmp.c b/kernel/kcmp.c
> > index 5353edfad8e1..4fb23f242e0f 100644
> > --- a/kernel/kcmp.c
> > +++ b/kernel/kcmp.c
> > @@ -63,6 +63,9 @@ get_file_raw_ptr(struct task_struct *task, unsigned int idx)
> >  {
> >       struct file *file;
> >
> > +     /* This RCU locking is only present to silence warnings.  The pointer
> > +      * value is only used for comparison and not dereferenced, so it is
> > +      * acceptable. */
> >       rcu_read_lock();
> >       file = task_lookup_fd_rcu(task, idx);
> >       rcu_read_unlock();
>
> They are not wrong, this is just such a bit weird semantics where
> we fetch the pointers and strictly speaking map them into numbers
> set to compare. But I agree that such tricks might confuse. How about
>
>         /*
>          * Fetching file pointers inside RCU read-lock section
>          * and reuse them as plain numbers is done in a sake
>          * of speed. But make sure never dereference them after.
>          */

I would tweak it a little to "Fetch file pointers inside RCU read-lock
section, but skip additional locking for speed.  The pointer values
will be used as integers, and must not be dereferenced."

One other idea I had was to switch the return value to "void *".  That
way it isn't a struct file, and it isn't readily dereference-able.
But I wasn't sure if that would be overkill.  What do you think?

Thanks,
Jason