On Thu, Sep 03, 2020 at 08:57:01PM +0900, Tetsuo Handa wrote:
> On 2020/09/03 20:31, Wouter Verhelst wrote:
> > So.
> >
> > On Wed, Sep 02, 2020 at 08:09:54PM +0900, Tetsuo Handa wrote:
> >> syzbot is reporting hung task at nbd_ioctl() [1], for there are two
> >> problems regarding TIPC's connectionless socket's shutdown() operation.
> >> I found C reproducer for this problem (shown below) from "no output from
> >> test machine (2)" report.
> >>
> >> ----------
> >>
> >> int main(int argc, char *argv[])
> >> {
> >> const int fd = open("/dev/nbd0", 3);
> >> ioctl(fd, NBD_SET_SOCK, socket(PF_TIPC, SOCK_DGRAM, 0));
> >
> > NBD expects a stream socket, not a datagram one.
> >
> >> ioctl(fd, NBD_DO_IT, 0);
> >
> > This is supposed to sit and wait until someone disconnects the device
> > again (which you probably cna't do with datagram sockets). Changing that
> > changes a userspace API.
> >
>
> Excuse me, but other datagram sockets (e.g. socket(PF_INET, SOCK_DGRAM, 0))
> does not
> hit this problem. What do you want to do? Add a "whether the file descriptor
> passed
> to ioctl(NBD_SET_SOCK) is a SOCK_STREAM socket" test to the NBD side?
I missed originally that you were checking whether the passed socket is
in fact a SOCK_DGRAM socket, and limiting the changes to that. That's
fine, because NBD doesn't deal with SOCK_DGRAM sockets anyway (i.e.,
passing a SOCK_DGRAM socket to the NBD device is undefined behavior). If
the behavior also changes for SOCK_STREAM sockets then that would be a
problem that would need to be reverted, but otherwise it's fine.
--
To the thief who stole my anti-depressants: I hope you're happy
-- seen somewhere on the Internet on a photo of a billboard
