Re: net/sched: GPF in qdisc_hash_add

[Eric Dumazet]

On Thu, Mar 23, 2017 at 12:06 PM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>
> On Thu, Mar 23, 2017 at 8:00 PM, Cong Wang <xiyou.wangcong@xxxxxxxxx> wrote:
> > On Thu, Mar 23, 2017 at 9:06 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> >> kasan: CONFIG_KASAN_INLINE enabled
> >> kasan: GPF could be caused by NULL-ptr deref or user memory access
> >> general protection fault: 0000 [#1] SMP KASAN
> >> Dumping ftrace buffer:
> >>    (ftrace buffer empty)
> >> Modules linked in:
> >> CPU: 2 PID: 12732 Comm: syz-executor6 Not tainted 4.11.0-rc3+ #365
> >> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
> >> task: ffff880062b7a2c0 task.stack: ffff880033480000
> >> RIP: 0010:qdisc_hash_add.part.19+0xb6/0x3c0 net/sched/sch_api.c:280
> >> RSP: 0018:ffff880033487820 EFLAGS: 00010202
> >> RAX: dffffc0000000000 RBX: ffffffff85357e00 RCX: ffffc90002b24000
> >> RDX: 000000000000007a RSI: ffffffff835a523a RDI: 00000000000003d0
> >> RBP: ffff8800334878b8 R08: fffffbfff0a6afeb R09: fffffbfff0a6afeb
> >> R10: 0000000000000001 R11: fffffbfff0a6afea R12: ffffffff85357e48
> >> R13: 1ffff10006690f06 R14: ffff880033487890 R15: 0000000000000000
> >> FS:  00007f68665d0700(0000) GS:ffff88006e200000(0000) knlGS:0000000000000000
> >> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> >> CR2: 00000000004c2d44 CR3: 000000003c6f8000 CR4: 00000000000026e0
> >> Call Trace:
> >>  qdisc_hash_add+0x76/0x90 net/sched/sch_api.c:279
> >>  attach_default_qdiscs net/sched/sch_generic.c:798 [inline]
> >>  dev_activate+0x6ca/0x920 net/sched/sch_generic.c:829
> >>  __dev_open+0x25b/0x360 net/core/dev.c:1348
> >>  __dev_change_flags+0x159/0x3d0 net/core/dev.c:6460
> >>  dev_change_flags+0x88/0x140 net/core/dev.c:6525
> >>  dev_ifsioc+0x51f/0x9b0 net/core/dev_ioctl.c:254
> >>  dev_ioctl+0x1fe/0x1030 net/core/dev_ioctl.c:532
> >>  sock_do_ioctl+0x94/0xb0 net/socket.c:902
> >>  sock_ioctl+0x2c2/0x440 net/socket.c:993
> >>  vfs_ioctl fs/ioctl.c:45 [inline]
> >>  do_vfs_ioctl+0x1af/0x16d0 fs/ioctl.c:685
> >>  SYSC_ioctl fs/ioctl.c:700 [inline]
> >>  SyS_ioctl+0x8f/0xc0 fs/ioctl.c:691
> >>  entry_SYSCALL_64_fastpath+0x1f/0xc2
> >
> > The interesting part is why the NULL dereference is in
> > qdisc_hash_add(), since we have a check before calling
> > it:
> >
> > #ifdef CONFIG_NET_SCHED
> >         if (dev->qdisc)
> >                 qdisc_hash_add(dev->qdisc);
> > #endif
> >
> >
> > When attach_one_default_qdisc() fails, we should trigger
> > the NULL pointer dereference bug at:
> >
> >     atomic_inc(&dev->qdisc->refcnt);
>
> I think qdisc is not NULL, it's something _in_ qdisc that is NULL. The
> crash happens here:
>
>                 struct Qdisc *root = qdisc_dev(q)->qdisc;
>
> so it's probably device.



Looks like this bug came with commit 59cc1f61f09c
("net: sched: convert qdisc linked list to hashtable")

I would simply guard qdisc_hash_add()

(Against &noop_qdisc)

diff --git a/net/sched/sch_api.c b/net/sched/sch_api.c
index bcf49cd2278670197f2a7e9d4e9a62ae8d117468..2bb34b51cffdb05434a488b9f45c344d57868253
100644
--- a/net/sched/sch_api.c
+++ b/net/sched/sch_api.c
@@ -276,6 +276,8 @@ static struct Qdisc *qdisc_match_from_root(struct
Qdisc *root, u32 handle)

 void qdisc_hash_add(struct Qdisc *q)
 {
+ if (q == &noop_qdisc)
+ return;
  if ((q->parent != TC_H_ROOT) && !(q->flags & TCQ_F_INGRESS)) {
  struct Qdisc *root = qdisc_dev(q)->qdisc;