net_sched: fix a refcount_t issue with noop_qdisc
authorEric Dumazet <edumazet@google.com>
Fri, 25 Aug 2017 04:12:28 +0000 (21:12 -0700)
committerDavid S. Miller <davem@davemloft.net>
Fri, 25 Aug 2017 04:28:24 +0000 (21:28 -0700)
syzkaller reported a refcount_t warning [1]

Issue here is that noop_qdisc refcnt was never really considered as
a true refcount, since qdisc_destroy() found TCQ_F_BUILTIN set :

if (qdisc->flags & TCQ_F_BUILTIN ||
    !refcount_dec_and_test(&qdisc->refcnt)))
return;

Meaning that all atomic_inc() we did on noop_qdisc.refcnt were not
really needed, but harmless until refcount_t came.

To fix this problem, we simply need to not increment noop_qdisc.refcnt,
since we never decrement it.

[1]
refcount_t: increment on 0; use-after-free.
------------[ cut here ]------------
WARNING: CPU: 0 PID: 21754 at lib/refcount.c:152 refcount_inc+0x47/0x50 lib/refcount.c:152
Kernel panic - not syncing: panic_on_warn set ...

CPU: 0 PID: 21754 Comm: syz-executor7 Not tainted 4.13.0-rc6+ #20
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Call Trace:
 __dump_stack lib/dump_stack.c:16 [inline]
 dump_stack+0x194/0x257 lib/dump_stack.c:52
 panic+0x1e4/0x417 kernel/panic.c:180
 __warn+0x1c4/0x1d9 kernel/panic.c:541
 report_bug+0x211/0x2d0 lib/bug.c:183
 fixup_bug+0x40/0x90 arch/x86/kernel/traps.c:190
 do_trap_no_signal arch/x86/kernel/traps.c:224 [inline]
 do_trap+0x260/0x390 arch/x86/kernel/traps.c:273
 do_error_trap+0x120/0x390 arch/x86/kernel/traps.c:310
 do_invalid_op+0x1b/0x20 arch/x86/kernel/traps.c:323
 invalid_op+0x1e/0x30 arch/x86/entry/entry_64.S:846
RIP: 0010:refcount_inc+0x47/0x50 lib/refcount.c:152
RSP: 0018:ffff8801c43477a0 EFLAGS: 00010282
RAX: 000000000000002b RBX: ffffffff86093c14 RCX: 0000000000000000
RDX: 000000000000002b RSI: ffffffff8159314e RDI: ffffed0038868ee8
RBP: ffff8801c43477a8 R08: 0000000000000001 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000000 R12: ffffffff86093ac0
R13: 0000000000000001 R14: ffff8801d0f3bac0 R15: dffffc0000000000
 attach_default_qdiscs net/sched/sch_generic.c:792 [inline]
 dev_activate+0x7d3/0xaa0 net/sched/sch_generic.c:833
 __dev_open+0x227/0x330 net/core/dev.c:1380
 __dev_change_flags+0x695/0x990 net/core/dev.c:6726
 dev_change_flags+0x88/0x140 net/core/dev.c:6792
 dev_ifsioc+0x5a6/0x930 net/core/dev_ioctl.c:256
 dev_ioctl+0x2bc/0xf90 net/core/dev_ioctl.c:554
 sock_do_ioctl+0x94/0xb0 net/socket.c:968
 sock_ioctl+0x2c2/0x440 net/socket.c:1058
 vfs_ioctl fs/ioctl.c:45 [inline]
 do_vfs_ioctl+0x1b1/0x1520 fs/ioctl.c:685
 SYSC_ioctl fs/ioctl.c:700 [inline]
 SyS_ioctl+0x8f/0xc0 fs/ioctl.c:691

Fixes: 7b9364050246 ("net, sched: convert Qdisc.refcnt from atomic_t to refcount_t")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Cc: Reshetova, Elena <elena.reshetova@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/sch_generic.h
net/sched/sch_api.c
net/sched/sch_generic.c

index 67f815e5d52517390226bc3531b1ea7b5f1020bc..c1109cdbbfa6afb9aff0d6033aef7b615630ffc1 100644 (file)
@@ -101,6 +101,13 @@ struct Qdisc {
        spinlock_t              busylock ____cacheline_aligned_in_smp;
 };
 
+static inline void qdisc_refcount_inc(struct Qdisc *qdisc)
+{
+       if (qdisc->flags & TCQ_F_BUILTIN)
+               return;
+       refcount_inc(&qdisc->refcnt);
+}
+
 static inline bool qdisc_is_running(const struct Qdisc *qdisc)
 {
        return (raw_read_seqcount(&qdisc->running) & 1) ? true : false;
index a3fa144b864871088209386fd573bded1886432f..4fb5a3222d0d324167f079f755be14eb028b4a50 100644 (file)
@@ -836,7 +836,7 @@ static int qdisc_graft(struct net_device *dev, struct Qdisc *parent,
 
                        old = dev_graft_qdisc(dev_queue, new);
                        if (new && i > 0)
-                               refcount_inc(&new->refcnt);
+                               qdisc_refcount_inc(new);
 
                        if (!ingress)
                                qdisc_destroy(old);
@@ -847,7 +847,7 @@ skip:
                        notify_and_destroy(net, skb, n, classid,
                                           dev->qdisc, new);
                        if (new && !new->ops->attach)
-                               refcount_inc(&new->refcnt);
+                               qdisc_refcount_inc(new);
                        dev->qdisc = new ? : &noop_qdisc;
 
                        if (new && new->ops->attach)
@@ -1256,7 +1256,7 @@ replay:
                                if (q == p ||
                                    (p && check_loop(q, p, 0)))
                                        return -ELOOP;
-                               refcount_inc(&q->refcnt);
+                               qdisc_refcount_inc(q);
                                goto graft;
                        } else {
                                if (!q)
index 57ba406f1437323a4ba172d52f14a8b687b86708..4ba6da5fb2546c35ad48fe1f3632df8ca9957b34 100644 (file)
@@ -785,7 +785,7 @@ static void attach_default_qdiscs(struct net_device *dev)
            dev->priv_flags & IFF_NO_QUEUE) {
                netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
                dev->qdisc = txq->qdisc_sleeping;
-               refcount_inc(&dev->qdisc->refcnt);
+               qdisc_refcount_inc(dev->qdisc);
        } else {
                qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT);
                if (qdisc) {