Re: [PATCH] net/9p: Fix sparse rcu warnings in client.c

[Alexander Kapshuk]

On Thu, Jun 18, 2020 at 10:08 PM Dominique Martinet
<asmadeus@xxxxxxxxxxxxx> wrote:
>
> Alexander Kapshuk wrote on Thu, Jun 18, 2020:
> > Address sparse nonderef rcu warnings:
> > net/9p/client.c:790:17: warning: incorrect type in argument 1 (different address spaces)
> > net/9p/client.c:790:17:    expected struct spinlock [usertype] *lock
> > net/9p/client.c:790:17:    got struct spinlock [noderef] <asn:4> *
> > net/9p/client.c:792:48: warning: incorrect type in argument 1 (different address spaces)
> > net/9p/client.c:792:48:    expected struct spinlock [usertype] *lock
> > net/9p/client.c:792:48:    got struct spinlock [noderef] <asn:4> *
> > net/9p/client.c:872:17: warning: incorrect type in argument 1 (different address spaces)
> > net/9p/client.c:872:17:    expected struct spinlock [usertype] *lock
> > net/9p/client.c:872:17:    got struct spinlock [noderef] <asn:4> *
> > net/9p/client.c:874:48: warning: incorrect type in argument 1 (different address spaces)
> > net/9p/client.c:874:48:    expected struct spinlock [usertype] *lock
> > net/9p/client.c:874:48:    got struct spinlock [noderef] <asn:4> *
> >
> > Signed-off-by: Alexander Kapshuk <alexander.kapshuk@xxxxxxxxx>
>
> Thanks for this patch.
> From what I can see, there are tons of other parts of the code doing the
> same noderef access pattern to access current->sighand->siglock and I
> don't see much doing that.
> A couple of users justify this by saying SLAB_TYPESAFE_BY_RCU ensures
> we'll always get a usable lock which won't be reinitialized however we
> access it... It's a bit dubious we'll get the same lock than unlock to
> me, so I agree to some change though.
>
> After a second look I think we should use something like the following:
>
> if (!lock_task_sighand(current, &flags))
>         warn & skip (or some error, we'd null deref if this happened currently);
> recalc_sigpending();
> unlock_task_sighand(current, &flags);
>
> As you can see, the rcu_read_lock() isn't kept until the unlock so I'm
> not sure it will be enough to please sparse, but I've convinced myself
> current->sighand cannot change while we hold the lock and there just are
> too many such patterns in the kernel.
>
> Please let me know if I missed something or if there is an ongoing
> effort to change how this works; I'll wait for a v2.
>
> --
> Dominique

Thanks for your prompt response.
I too made the same observation of the numerous patterns in the kernel
where current->sighand is accessed without being rcu_dereference()'d.
For this patch I used kernel/signal.c:1368,1398: __lock_task_sighand()
as an example.

I will give your suggestion a careful consideration and will get back
to you soon.
Thanks.