Re: [PATCH 2/4] io_uring/cmd: Introduce SOCKET_URING_OP_GETSOCKOPT

[Breno Leitao <leitao@debian.org>]

On Fri, Jul 28, 2023 at 11:07:10AM -0700, Stanislav Fomichev wrote:
> On Fri, Jul 28, 2023 at 10:03 AM Breno Leitao <leitao@debian.org> wrote:
> >
> > Hello Stanislav,
> >
> > On Tue, Jul 25, 2023 at 10:02:40AM -0700, Stanislav Fomichev wrote:
> > > On 07/25, Breno Leitao wrote:
> > > > On Mon, Jul 24, 2023 at 10:31:28AM -0700, Stanislav Fomichev wrote:
> > > > > On 07/24, Breno Leitao wrote:
> > > > > > Add support for getsockopt command (SOCKET_URING_OP_GETSOCKOPT), where
> > > > > > level is SOL_SOCKET. This is leveraging the sockptr_t infrastructure,
> > > > > > where a sockptr_t is either userspace or kernel space, and handled as
> > > > > > such.
> > > > > >
> > > > > > Function io_uring_cmd_getsockopt() is inspired by __sys_getsockopt().
> > > > >
> > > > > We probably need to also have bpf bits in the new
> > > > > io_uring_cmd_getsockopt?
> > > >
> > > > It might be interesting to have the BPF hook for this function as
> > > > well, but I would like to do it in a following patch, so, I can
> > > > experiment with it better, if that is OK.
> >
> > I spent smoe time looking at the problem, and I understand we want to
> > call something as BPF_CGROUP_RUN_PROG_{G,S}ETSOCKOPT() into
> > io_uring_cmd_{g,s}etsockopt().
> >
> > Per the previous conversation with Williem,
> > io_uring_cmd_{g,s}etsockopt() should use optval as a user pointer (void __user
> > *optval), and optlen as a kernel integer (it comes as from the io_uring
> > SQE), such as:
> >
> >         void __user *optval = u64_to_user_ptr(READ_ONCE(cmd->sqe->optval));
> >         int optlen = READ_ONCE(cmd->sqe->optlen);
> >
> > Function BPF_CGROUP_RUN_PROG_GETSOCKOPT() calls
> > __cgroup_bpf_run_filter_getsockopt() which expects userpointer for
> > optlen and optval.
> >
> > At the same time BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN() expects kernel
> > pointers for both optlen and optval.
> >
> > In this current patchset, it has user pointer for optval and kernel value
> > for optlen. I.e., a third combination.  So, none of the functions would
> > work properly, and we probably do not want to create another function.
> >
> > I am wondering if it is a good idea to move
> > __cgroup_bpf_run_filter_getsockopt() to use sockptr_t, so, it will be
> > able to adapt to any combination.
> 
> Yeah, I think it makes sense. However, note that the intent of that
> optlen being a __user pointer is to possibly write some (updated)
> value back into the userspace.
> Presumably, you'll pass that updated optlen into some io_uring
> completion queue? (maybe a stupid question, not super familiar with
> io_uring)

On io_uring proposal, the optlen is part of the SQE for setsockopt().
You give a  userpointer (optval) and set the optlen in the SQE->optlen.

For getsockopt(), the optlen is returned as a result of the operation,
in the CQE->res.

If you need more detail about it, I documented this behaviour in the
cover-letter (PS1):

https://lore.kernel.org/all/20230724142237.358769-1-leitao@debian.org/

Thanks for the feedback!