Re: tty: n_gsm: race condition in gsmld_ioctl

[Jiri Slaby <jirislaby@kernel.org>]

Hello & thanks.

To: Daniel

On 16. 04. 24, 12:57, Yewon Choi wrote:
> Hello,
> 
> memory leak in gsmld_ioctl was reported by syzbot, and this issue seems to be
> raised from race condition involving gsm->dlci[addr].
> 
> https://syzkaller.appspot.com/bug?extid=6e3e8f30f269f5028e5d
> 
> When dlci object is checked and initialized in gsmld_ioctl(), it may be
> assigned multiple times because there is no lock which guards entering
> the critical section in gsm_dlci_alloc() from multiple threads.
> 
> For example, when multiple ioctl() are called concurrently, the following
> scenario is possible:
> 
> Thread 0                                Thread 1
> ioctl(GSMIOC_SETCONF_DLCI)              ioctl(GSMIOC_GETCONF_DLCI)
> 
> gsmld_ioctl():                          gsmld_ioctl():
>    dlci = gsm->dlci[addr];
>                                            dlci = gsm->dlci[addr];
>    if (!dlci) {
>      gsm_dlci_alloc(gsm, addr):
>        ...
>        gsm->dlci[addr] = dlci;
>                                            if (!dlci) {
>                                              gsm_dlci_alloc(gsm, addr):
>                                                ...
>                                                // overwritten & memory leak
>                                                gsm->dlci[addr] = dlci;
> 
> 
> We think either (1) gsm_dlci_alloc() should hold a lock(mutex) and do
> internal check about whether gsm->dlci[addr] is NUll or not, OR
> (2) all callers of gsm_dlci_alloc() should hold gsm->mutex and check
> whether gsm->dlci[addr] is NUll or not (like gsmtty_install()).
> 
> Could you check this? If it makes sense, we will write a patch following
> one of the suggestions.
> 
> Best Regards,
> Yewon Choi

-- 
js