On 2022/03/30 8:51, Dominique Martinet wrote:
> Tetsuo Handa wrote on Wed, Mar 30, 2022 at 07:35:47AM +0900:
>> This seems to be an example of
>> https://lkml.kernel.org/r/[email protected]
>> introduced by "ext4: truncate during setxattr leads to kernel panic".
>
> Thanks for the pointer
>
>> Please don't use schedule_work() if you need to use flush_scheduled_work().
>
> In this case we don't call flush_scheduled_work -- ext4 does.
Yes, that's why I changed recipients to ext4 people.
> The problem is mixing in the two subsystems when someone (e.g. syzbot)
> opens an ext4 file and passes that fd to 9p when mounting with e.g.
> mount -t 9p -o rfdno=<no>,wfdno=<no>
>
> Frankly that's just not something I consider useful, interacting through
> 9p to a local file doesn't make sense except for testing.
>
> If that is a real problem, the simplest way out would be to just forbid
> non-socket FDs if it's something we can check.
Do you mean that p9_fd_open() in net/9p/trans_fd.c does not need to accept non-socket file descriptors?
Then, it's something you can check. You can use S_ISSOCK() like e.g. netlink_getsockbyfilp() does.
Tetsuo Handa wrote on Wed, Mar 30, 2022 at 10:57:15AM +0900:
> >> Please don't use schedule_work() if you need to use flush_scheduled_work().
> >
> > In this case we don't call flush_scheduled_work -- ext4 does.
>
> Yes, that's why I changed recipients to ext4 people.
Sorry, I hadn't noticed.
9p is the one calling schedule_work, so ultimately it really is the
combinaison of the two, and not just ext4 that's wrong here.
> > The problem is mixing in the two subsystems when someone (e.g. syzbot)
> > opens an ext4 file and passes that fd to 9p when mounting with e.g.
> > mount -t 9p -o rfdno=<no>,wfdno=<no>
> >
> > Frankly that's just not something I consider useful, interacting through
> > 9p to a local file doesn't make sense except for testing.
> >
> > If that is a real problem, the simplest way out would be to just forbid
> > non-socket FDs if it's something we can check.
>
> Do you mean that p9_fd_open() in net/9p/trans_fd.c does not need to
> accept non-socket file descriptors?
Yes, I can't think of any valid usage that would involve non-socket fd
there.
It might be useful to leave as a test vector, but if it causes problems
I think it's perfectly OK to just refuse these.
> Then, it's something you can check. You can use S_ISSOCK() like
> e.g. netlink_getsockbyfilp() does
Thanks for the example
--
Dominique