Data race on dev->mtu betwen __dev_set_mtu() and rawv6_send_hdrinc()

Data race on dev->mtu betwen __dev_set_mtu() and rawv6_send_hdrinc()

[Gong, Sishuai]

Hello,

We found a data race on dev->mtu between function __dev_set_mtu() and rawv6_send_hdrinc(). It happens with the following interleaving. 

writer: __dev_set_mtu()										reader: rawv6_send_hdrinc()
														if (length > rt->dst.dev->mtu) {
	WRITE_ONCE(dev->mtu, new_mtu);
															ipv6_local_error(sk, EMSGSIZE, fl6, rt->dst.dev->mtu);

If the writer happens to change dev->mtu to a value that is bigger than the variable ‘length’, then ipv6_local_error will read a value that doesn’t satisfy this conditional statement. While there is no need to use lock to protect the read, it is probably better to only read dev->mtu once in rawv6_send_hdrinc().


Thanks,
Sishuai

Re: Data race on dev->mtu betwen __dev_set_mtu() and rawv6_send_hdrinc()

[Willem de Bruijn]

On Mon, Feb 22, 2021 at 6:30 PM Gong, Sishuai <sishuai@purdue.edu> wrote:
>
> Hello,
>
> We found a data race on dev->mtu between function __dev_set_mtu() and rawv6_send_hdrinc(). It happens with the following interleaving.
>
> writer: __dev_set_mtu()                                                                         reader: rawv6_send_hdrinc()
>                                                                                                                 if (length > rt->dst.dev->mtu) {
>         WRITE_ONCE(dev->mtu, new_mtu);
>                                                                                                                         ipv6_local_error(sk, EMSGSIZE, fl6, rt->dst.dev->mtu);
>
> If the writer happens to change dev->mtu to a value that is bigger than the variable ‘length’, then ipv6_local_error will read a value that doesn’t satisfy this conditional statement. While there is no need to use lock to protect the read, it is probably better to only read dev->mtu once in rawv6_send_hdrinc().

Makes sense. The same would then apply to raw_send_hdrinc().