Re: [PATCH net-next 1/3] net: rework SIOCGSTAMP ioctl handling

[Willem de Bruijn]

On Wed, Apr 17, 2019 at 12:19 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
>
> On Wed, Apr 17, 2019 at 4:46 PM Willem de Bruijn
> <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > On Tue, Apr 16, 2019 at 4:38 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > > The SIOCGSTAMP/SIOCGSTAMPNS ioctl commands are implemented by many
> > > socket protocol handlers, and all of those end up calling the same
> > > sock_get_timestamp()/sock_get_timestampns() helper functions, which
> > > results in a lot of duplicate code.
> > >
> > > With the introduction of 64-bit time_t on 32-bit architectures, this
> > > gets worse, as we then need four different ioctl commands in each
> > > socket protocol implementation.
> > >
> > > To simplify that, let's add a new .gettstamp() operation in
> > > struct proto_ops, and move ioctl implementation into the common
> > > sock_ioctl()/compat_sock_ioctl_trans() functions that these all go
> > > through.
> > >
> > > We can reuse the sock_get_timestamp() implementation, but generalize
> > > it so it can deal with both native and compat mode, as well as
> > > timeval and timespec structures.
> > >
> > > Acked-by: Stefan Schmidt <stefan@xxxxxxxxxxxxxxxxxx>
> > > Link: https://lore.kernel.org/lkml/CAK8P3a038aDQQotzua_QtKGhq8O9n+rdiz2=WDCp82ys8eUT+A@xxxxxxxxxxxxxx/
> > > Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> > > ---
> > > v2: reworked to not break sparc64 support
> >
> > From the discussion of v1 I thought you planned to unconditionally
> > call sock_gettstamp() for all protocols, avoiding the need to plumb in
> > all these new callbacks?
> >
> > That is more concise, though this closer to the existing behavior. So,
> > fine either way.
>
> Thanks for the reminder. I have definitely waited too long before revisiting
> this series, and only had a vague recollection of that discussion but could
> not find it in the logs (found it now, and the Link I quoted...).
>
> I would prefer to get this series into the coming merge window, and
> probably don't have time to rework it completely by then, so I hope
> the current version is ok.

Absolutely. This is a great simplification either way.

> I also found your comment on lock_sock(), which could be easily
> added inside of sock_gettstamp() if you think we should have that.

To remind, the issue is that sock_enable_timestamp should update both
sk_flags and net_enable_timestamp as one atomic operation, by holding
the socket lock. The lock is held when called from a setsockopt path.
And from some ioctl() implementations. The syzkaller reproducer
triggered through inet_release, so the most widely used paths are
buggy today.

Since the current state is inconsistent already, we can defer the fix.
I expect that it is now safe to add a lock around this logic without
triggering lockdep issues in any of the numerous paths. But just in
case, it still seems safer to do that in a separate patch that we can
revert or rework independent from this, if that would prove necessary.

> There is one more issue I just noticed (I dropped the necessary
> sock_read_timestamp()), so I have to repost the series anyway
> to fix that.

Instead of reading sk->sk_stamp directly? Yes, makes sense.