[PROBLEM] a data race between tcp_set_default_congestion_control() and tcp_set_congestion_control()

[PROBLEM] a data race between tcp_set_default_congestion_control() and tcp_set_congestion_control()

[Gong, Sishuai]

Hi,

We found a data race between tcp_set_default_congestion_control() and tcp_set_congestion_control() in linux-5.12-rc3. 
In general, when tcp_set_congestion_control() is reading ca->flags with a lock grabbed, tcp_set_default_congestion_control() 
may be updating ca->flags at the same time, as shown below.

When the writer and reader are running parallel, tcp_set_congestion_control()’s control flow 
might be non-deterministic, either returning a -EPERM or calling tcp_reinit_congestion_control().

We also notice in tcp_set_allowed_congestion_control(), the write to ca->flags is protected by tcp_cong_list_lock,
so we want to point it out in case the data race is unexpected.

Thread 1							Thread 2
//tcp_set_default_congestion_control()	//tcp_set_congestion_control()
								// lock_sock() grabbed
								if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
									err = -EPERM;
								else if (!bpf_try_module_get(ca, ca->owner))
									err = -EBUSY;
								else
									tcp_reinit_congestion_control(sk, ca);
ca->flags |= TCP_CONG_NON_RESTRICTED;
	


Thanks,
Sishuai

Re: [PROBLEM] a data race between tcp_set_default_congestion_control() and tcp_set_congestion_control()

[Eric Dumazet]

On Thu, Apr 15, 2021 at 5:47 PM Gong, Sishuai <sishuai@purdue.edu> wrote:
>
> Hi,
>
> We found a data race between tcp_set_default_congestion_control() and tcp_set_congestion_control() in linux-5.12-rc3.
> In general, when tcp_set_congestion_control() is reading ca->flags with a lock grabbed, tcp_set_default_congestion_control()
> may be updating ca->flags at the same time, as shown below.
>
> When the writer and reader are running parallel, tcp_set_congestion_control()’s control flow
> might be non-deterministic, either returning a -EPERM or calling tcp_reinit_congestion_control().
>
> We also notice in tcp_set_allowed_congestion_control(), the write to ca->flags is protected by tcp_cong_list_lock,
> so we want to point it out in case the data race is unexpected.
>
> Thread 1                                                        Thread 2
> //tcp_set_default_congestion_control()  //tcp_set_congestion_control()
>                                                                 // lock_sock() grabbed
>                                                                 if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
>                                                                         err = -EPERM;
>                                                                 else if (!bpf_try_module_get(ca, ca->owner))
>                                                                         err = -EBUSY;
>                                                                 else
>                                                                         tcp_reinit_congestion_control(sk, ca);
> ca->flags |= TCP_CONG_NON_RESTRICTED;
>
>
>
> Thanks,
> Sishuai
>

Yes, obviously reading ca->flags while another thread might set the bit is racy.

This is of no consequence, if you want to silence KCSAN please a patch.