[PATCH] net/socket: Acquire cgroup_lock in do_sock_getsockopt

[Eric Dumazet]

On Mon, Aug 19, 2024 at 10:27 AM Tze-nan Wu <Tze-nan.Wu at mediatek.com> wrote:
>
> The return value from `cgroup_bpf_enabled(CGROUP_GETSOCKOPT)` can change
> between the invocations of `BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN` and
> `BPF_CGROUP_RUN_PROG_GETSOCKOPT`.
>
> If `cgroup_bpf_enabled(CGROUP_GETSOCKOPT)` changes from "false" to
> "true"
> between the invocations of `BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN` and
> `BPF_CGROUP_RUN_PROG_GETSOCKOPT`,
> `BPF_CGROUP_RUN_PROG_GETSOCKOPT` will receive an -EFAULT from
> `__cgroup_bpf_run_filter_getsockopt(max_optlen=0)` due to `get_user()`
> had not reached in `BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN`.
>
> Scenario shown as below:
>
>            `process A`                      `process B`
>            -----------                      ------------
>   BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN
>                                             enable CGROUP_GETSOCKOPT
>   BPF_CGROUP_RUN_PROG_GETSOCKOPT (-EFAULT)
>
> Prevent `cgroup_bpf_enabled(CGROUP_GETSOCKOPT)` change between
> `BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN` and `BPF_CGROUP_RUN_PROG_GETSOCKOPT`
> by acquiring cgroup_lock.
>
> Co-developed-by: Yanghui Li <yanghui.li at mediatek.com>
> Signed-off-by: Yanghui Li <yanghui.li at mediatek.com>
> Co-developed-by: Cheng-Jui Wang <cheng-jui.wang at mediatek.com>
> Signed-off-by: Cheng-Jui Wang <cheng-jui.wang at mediatek.com>
> Signed-off-by: Tze-nan Wu <Tze-nan.Wu at mediatek.com>
>
> ---
>
> We have encountered this issue by observing that process A could sometimes
> get an -EFAULT from getsockopt() during our device boot-up, while another
> process B triggers the race condition by enabling CGROUP_GETSOCKOPT
> through bpf syscall at the same time.
>
> The race condition is shown below:
>
>            `process A`                        `process B`
>            -----------                        ------------
>   BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN
>
>                                               bpf syscall
>                                         (CGROUP_GETSOCKOPT enabled)
>
>   BPF_CGROUP_RUN_PROG_GETSOCKOPT
>   -> __cgroup_bpf_run_filter_getsockopt
>     (-EFAULT)
>
> __cgroup_bpf_run_filter_getsockopt return -EFAULT at the line shown below:
>         if (optval && (ctx.optlen > max_optlen || ctx.optlen < 0)) {
>                 if (orig_optlen > PAGE_SIZE && ctx.optlen >= 0) {
>                         pr_info_once("bpf getsockopt: ignoring program buffer with optlen=%d (max_optlen=%d)\n",
>                                      ctx.optlen, max_optlen);
>                         ret = retval;
>                         goto out;
>                 }
>                 ret = -EFAULT; <== return EFAULT here
>                 goto out;
>         }
>
> This patch should fix the race but not sure if it introduces any potential
> side effects or regression.
>
> And we wondering if this is a real issue in do_sock_getsockopt or if
> getsockopt() is designed to expect such race conditions.
> Should the userspace caller always anticipate an -EFAULT from getsockopt()
> if another process enables CGROUP_GETSOCKOPT at the same time?
>
> Any comment will be appreciated!
>
> BTW, I added Chengjui and Yanghui to Co-developed due to we have several
> discussions on this issue. And we both spend some time on this issue.
>
> ---
>  net/socket.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/net/socket.c b/net/socket.c
> index fcbdd5bc47ac..e0b2b16fd238 100644
> --- a/net/socket.c
> +++ b/net/socket.c
> @@ -2370,8 +2370,10 @@ int do_sock_getsockopt(struct socket *sock, bool compat, int level,
>         if (err)
>                 return err;
>
> -       if (!compat)
> +       if (!compat) {
> +               cgroup_lock();
>                 max_optlen = BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN(optlen);
> +       }
>

Acquiring cgroup_lock mutex in socket getsockopt() fast path ?

There is no way we can accept such a patch, please come up with a
reasonable patch.

cgroup_bpf_enabled() should probably be used once.