Re: general protection fault in do_move_mount (2)

From: Dmitry Vyukov
Date: Mon Jul 01 2019 - 10:59:19 EST


On Sat, Jun 29, 2019 at 10:39 PM Eric Biggers <ebiggers@xxxxxxxxxx> wrote:
>
> On Mon, Jun 24, 2019 at 11:28:18AM +0200, 'Dmitry Vyukov' via syzkaller-bugs wrote:
> > On Tue, Jun 18, 2019 at 4:03 PM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, Jun 18, 2019 at 03:47:10AM -0700, syzbot wrote:
> > > > Hello,
> > > >
> > > > syzbot found the following crash on:
> > > >
> > > > HEAD commit: 9e0babf2 Linux 5.2-rc5
> > > > git tree: upstream
> > > > console output: https://syzkaller.appspot.com/x/log.txt?x=138b310aa00000
> > > > kernel config: https://syzkaller.appspot.com/x/.config?x=d16883d6c7f0d717
> > > > dashboard link: https://syzkaller.appspot.com/bug?extid=6004acbaa1893ad013f0
> > > > compiler: gcc (GCC) 9.0.0 20181231 (experimental)
> > > > syz repro: https://syzkaller.appspot.com/x/repro.syz?x=154e8c2aa00000
> > >
> > > IDGI...
> > >
> > > mkdir(&(0x7f0000632000)='./file0\x00', 0x0)
> > > mount(0x0, 0x0, 0x0, 0x0, 0x0)
> > > syz_open_procfs(0x0, 0x0)
> > > r0 = open(&(0x7f0000032ff8)='./file0\x00', 0x0, 0x0)
> > > r1 = memfd_create(&(0x7f00000001c0)='\xb3', 0x0)
> > > write$FUSE_DIRENT(r1, &(0x7f0000000080)=ANY=[], 0x29)
> > > move_mount(r0, &(0x7f0000000040)='./file0\x00', 0xffffffffffffff9c, &(0x7f0000000100)='./file0\x00', 0x66)
> > >
> > > reads as if we'd done mkdir ./file0, opened it and then tried
> > > to feed move_mount(2) "./file0" relative to that descriptor.
> > > How the hell has that avoided an instant -ENOENT? On the first
> > > pair, that is - the second one (AT_FDCWD, "./file0") is fine...
> > >
> > > Confused... Incidentally, what the hell is
> > > mount(0x0, 0x0, 0x0, 0x0, 0x0)
> > > about? *IF* that really refers to mount(2) with
> > > such arguments, all you'll get is -EFAULT. Way before
> > > it gets to actually doing anything - it won't get past
> > > /* ... and get the mountpoint */
> > > retval = user_path(dir_name, &path);
> > > if (retval)
> > > return retval;
> > > in do_mount(2)...
> >
> > Yes, mount(0x0, 0x0, 0x0, 0x0, 0x0) is mount with 0 arguments. Most
> > likely it returns EFAULT.
> > Since the reproducer have "threaded":true,"collide":true and no C
> > repro, most likely this is a subtle race. So it attempted to remove
> > mount(0x0, 0x0, 0x0, 0x0, 0x0) but it did not crash, so the conclusion
> > was that it's somehow needed. You can actually see that other
> > reproducers for this bug do not have this mount, but are otherwise
> > similar.
> >
> > With "threaded":true,"collide":true the execution mode is not just
> > "execute each syscall once one after another".
> > The syscalls are executed in separate threads and actually twice. You
> > can see the approximate execution mode in this C program:
> > https://gist.githubusercontent.com/dvyukov/c3a52f012e7cff9bdebf3935d35245cf/raw/b5587824111a1d982c985b00137ae8609572335b/gistfile1.txt
> > Yet using the C program did not trigger the crash somehow (maybe just
> > slightly different timings).
> >
> > Since syzkaller was able to reproduce it multiple times, it looks like
> > a real bug rather than an induced memory corruption or something.
> >
>
> I sent a patch to fix this bug (https://lore.kernel.org/linux-fsdevel/20190629202744.12396-1-ebiggers@xxxxxxxxxx/T/#u)
>
> Dmitry, any idea why syzbot found such a bizarre reproducer for this?
> This is actually reproducible by a simple single threaded program:
>
> #include <unistd.h>
>
> #define __NR_move_mount 429
> #define MOVE_MOUNT_F_EMPTY_PATH 0x00000004
>
> int main()
> {
> int fds[2];
>
> pipe(fds);
> syscall(__NR_move_mount, fds[0], "", -1, "/", MOVE_MOUNT_F_EMPTY_PATH);
> }


There is no pipe in the reproducer, so it could not theoretically come
up with the reproducer with the pipe. During minimization syzkaller
only tries to remove syscalls and simplify arguments and execution
mode.
What would be the simplest reproducer expressed as further
minimization of this reproducer?
https://syzkaller.appspot.com/x/repro.syz?x=154e8c2aa00000
I assume one of the syscalls is still move_mount, but what is the
other one? If it's memfd_create, or open of the procfs file, then it
seems that [ab]used heavy threading and syscall colliding as way to do
an arbitrary mutation of the program. Per se results of
memfd_create/procfs are not passed to move_mount. But by abusing races
it probably managed to do so in small percent of cases. It would also
explain why it's hard to reproduce.


> FYI, it also isn't really appropriate for syzbot to bisect all bugs in new
> syscalls to wiring them up to x86, and then blame all the x86 maintainers.
> Normally such bugs will be in the syscall itself, regardless of architecture.

Agree. Do you think it's something worth handling automatically
(stands out of the long tail of other inappropriate cases)? If so, how
could we detect such cases? It seems that some of these predicates are
quite hard to program. Similar things happen with introduction of new
bug detection tools and checks, wiring any functionality to new access
points and similar things.