general protection fault in sockfs_setattr

[shankarapailoor]

Hi,

I have been fuzzing Linux 4.17-rc7 with Syzkaller and found the
following crash: https://pastebin.com/ixX3RB9j

Syzkaller isolated the cause of the bug to the following program:

socketpair$unix(0x1, 0x1, 0x0,
&(0x7f0000000000)={<r0=>0xffffffffffffffff, <r1=>0xffffffffffffffff})
getresuid(&(0x7f0000000080)=<r2=>0x0, &(0x7f00000000c0),
&(0x7f0000000700))r3 = getegid()
fchownat(r0, &(0x7f0000000040)='\x00', r2, r3, 0x1000)
dup3(r1, r0, 0x80000)


The problematic area appears to be here:

static int sockfs_setattr(struct dentry *dentry, struct iattr *iattr)
{
int err = simple_setattr(dentry, iattr);

if (!err && (iattr->ia_valid & ATTR_UID)) {
struct socket *sock = SOCKET_I(d_inode(dentry));

sock->sk->sk_uid = iattr->ia_uid; //KASAN GPF
}
return err;
}

If dup3 is called concurrently with fchownat then can sock->sk be NULL?

--
Regards,
Shankara Pailoor

[Cong Wang]

Although dup3() implies a close(), fd is refcnt'ted, if dup3() runs
concurrently with fchownat() it should not be closed until whoever
the last closes it.

Or maybe fchownat() doesn't even hold refcnt of fd, since it aims
to change the file backed.


Not sure if the following is sufficient, inode might need to be protected
with some lock...

diff --git a/net/socket.c b/net/socket.c
index f10f1d947c78..6294b4b3132e 100644
--- a/net/socket.c
+++ b/net/socket.c
@@ -537,7 +537,10 @@ static int sockfs_setattr(struct dentry *dentry,
struct iattr *iattr)

if (!err && (iattr->ia_valid & ATTR_UID)) {
struct socket *sock = SOCKET_I(d_inode(dentry));

- sock->sk->sk_uid = iattr->ia_uid;
+ if (sock->sk)
+ sock->sk->sk_uid = iattr->ia_uid;
+ else
+ err = -ENOENT;
}

return err;

[shankarapailoor]

Hi Cong,

I added that check and it seems to stop the crash. Like you said, I
don't see where the reference count for the file is increased. The
inode lock also seems to be held during this call.

Regards,
Shankara

--
Regards,
Shankara Pailoor

[Tetsuo Handa]

Pastebin says that it was 4.17.0-rc4+ rather than 4.17-rc7.
I suggest reporting to Al Viro and linux-fsdevel ML after
confirming that this bug still happens with linux.git , in
case this is a dentry related bug (e.g. someone is by error
calling dput() without getting a refcount).

Also, please don't eliminate kernel messages prior to the
crash. Sometimes previous kernel messages (e.g. memory
allocation fault injection) as-is indicate the cause.

[shankarapailoor]

++Al Viro

>Pastebin says that it was 4.17.0-rc4+ rather than 4.17-rc7.

Oops. Apologies Tetsuo. I confirmed the bug still happens with
linux.git. Here are the Syzkaller logs around the crash along with my
kernel configs.

Syzkaller Logs: https://pastebin.com/qqQyX0Ms
Config: https://pastebin.com/aEDARPDJ

Regards,
Shankara

--
Regards,
Shankara Pailoor

[Cong Wang]

On Tue, Jun 5, 2018 at 7:19 PM, shankarapailoor
<shankar...@gmail.com> wrote:
> Hi Cong,
>
> I added that check and it seems to stop the crash. Like you said, I
> don't see where the reference count for the file is increased. The
> inode lock also seems to be held during this call.

I know inode lock is held for ->setattr(), but not for ->release(), this
is why I suspect sock_close() could still race with sockfs_setattr()
after my patch.

I am not sure if it is crazy to just hold fd refcnt for fchmodat() too..