Re: [PATCH] net: check the return value of the copy_from_sockptr

[Stanislav Fomichev <stfomichev@gmail.com>]

On 09/11, Cong Wang wrote:
> On Wed, Sep 11, 2024 at 07:15:27PM +0200, Eric Dumazet wrote:
> > On Wed, Sep 11, 2024 at 6:58 PM Cong Wang <xiyou.wangcong@gmail.com> wrote:
> > >
> > > On Wed, Sep 11, 2024 at 11:12:24AM +0200, Eric Dumazet wrote:
> > > > On Wed, Sep 11, 2024 at 10:23 AM Qianqiang Liu <qianqiang.liu@163.com> wrote:
> > > > >
> > > > > > I do not think it matters, because the copy is performed later, with
> > > > > > all the needed checks.
> > > > >
> > > > > No, there is no checks at all.
> > > > >
> > > >
> > > > Please elaborate ?
> > > > Why should maintainers have to spend time to provide evidence to
> > > > support your claims ?
> > > > Have you thought about the (compat) case ?
> > > >
> > > > There are plenty of checks. They were there before Stanislav commit.
> > > >
> > > > Each getsockopt() handler must perform the same actions.
> > >
> > >
> > > But in line 2379 we have ops->getsockopt==NULL case:
> > >
> > > 2373         if (!compat)
> > > 2374                 copy_from_sockptr(&max_optlen, optlen, sizeof(int));
> > > 2375
> > > 2376         ops = READ_ONCE(sock->ops);
> > > 2377         if (level == SOL_SOCKET) {
> > > 2378                 err = sk_getsockopt(sock->sk, level, optname, optval, optlen);
> > > 2379         } else if (unlikely(!ops->getsockopt)) {
> > > 2380                 err = -EOPNOTSUPP;         // <--- HERE
> > > 2381         } else {
> > > 2382                 if (WARN_ONCE(optval.is_kernel || optlen.is_kernel,
> > > 2383                               "Invalid argument type"))
> > > 2384                         return -EOPNOTSUPP;
> > > 2385
> > > 2386                 err = ops->getsockopt(sock, level, optname, optval.user,
> > > 2387                                       optlen.user);
> > > 2388         }
> > >
> > > where we simply continue with calling BPF_CGROUP_RUN_PROG_GETSOCKOPT()
> > > which actually needs the 'max_optlen' we copied via copy_from_sockptr().
> > >
> > > Do I miss anything here?
> > 
> > This is another great reason why we should not change current behavior.
> 
> Hm? But the current behavior is buggy?
> 
> > 
> > err will be -EOPNOTSUPP, which was the original error code before
> > Stanislav patch.
> 
> You mean we should continue calling BPF_CGROUP_RUN_PROG_GETSOCKOPT()
> despite -EFAULT?
> 
> > 
> > Surely the eBPF program will use this value first, and not even look
> > at max_optlen
> > 
> > Returning -EFAULT might break some user programs, I don't know.
> 
> As you mentioned above, other ->getsockopt() already returns -EFAULT, so
> what is breaking? :)
> 
> > 
> > I feel we are making the kernel slower just because we can.
> 
> Safety and correctness also matter.

Can you explain what is not correct?

Calling BPF_CGROUP_RUN_PROG_GETSOCKOPT with max_optlen=0 should not be
a problem I think? (the buffer simply won't be accessible to the bpf prog)