1.. SPDX-License-Identifier: GPL-2.0
2
3=================
4Lockdep-RCU Splat
5=================
6
7Lockdep-RCU was added to the Linux kernel in early 2010
8(http://lwn.net/Articles/371986/).  This facility checks for some common
9misuses of the RCU API, most notably using one of the rcu_dereference()
10family to access an RCU-protected pointer without the proper protection.
11When such misuse is detected, an lockdep-RCU splat is emitted.
12
13The usual cause of a lockdep-RCU splat is someone accessing an
14RCU-protected data structure without either (1) being in the right kind of
15RCU read-side critical section or (2) holding the right update-side lock.
16This problem can therefore be serious: it might result in random memory
17overwriting or worse.  There can of course be false positives, this
18being the real world and all that.
19
20So let's look at an example RCU lockdep splat from 3.0-rc5, one that
21has long since been fixed::
22
23    =============================
24    WARNING: suspicious RCU usage
25    -----------------------------
26    block/cfq-iosched.c:2776 suspicious rcu_dereference_protected() usage!
27
28other info that might help us debug this::
29
30    rcu_scheduler_active = 1, debug_locks = 0
31    3 locks held by scsi_scan_6/1552:
32    #0:  (&shost->scan_mutex){+.+.}, at: [<ffffffff8145efca>]
33    scsi_scan_host_selected+0x5a/0x150
34    #1:  (&eq->sysfs_lock){+.+.}, at: [<ffffffff812a5032>]
35    elevator_exit+0x22/0x60
36    #2:  (&(&q->__queue_lock)->rlock){-.-.}, at: [<ffffffff812b6233>]
37    cfq_exit_queue+0x43/0x190
38
39    stack backtrace:
40    Pid: 1552, comm: scsi_scan_6 Not tainted 3.0.0-rc5 #17
41    Call Trace:
42    [<ffffffff810abb9b>] lockdep_rcu_dereference+0xbb/0xc0
43    [<ffffffff812b6139>] __cfq_exit_single_io_context+0xe9/0x120
44    [<ffffffff812b626c>] cfq_exit_queue+0x7c/0x190
45    [<ffffffff812a5046>] elevator_exit+0x36/0x60
46    [<ffffffff812a802a>] blk_cleanup_queue+0x4a/0x60
47    [<ffffffff8145cc09>] scsi_free_queue+0x9/0x10
48    [<ffffffff81460944>] __scsi_remove_device+0x84/0xd0
49    [<ffffffff8145dca3>] scsi_probe_and_add_lun+0x353/0xb10
50    [<ffffffff817da069>] ? error_exit+0x29/0xb0
51    [<ffffffff817d98ed>] ? _raw_spin_unlock_irqrestore+0x3d/0x80
52    [<ffffffff8145e722>] __scsi_scan_target+0x112/0x680
53    [<ffffffff812c690d>] ? trace_hardirqs_off_thunk+0x3a/0x3c
54    [<ffffffff817da069>] ? error_exit+0x29/0xb0
55    [<ffffffff812bcc60>] ? kobject_del+0x40/0x40
56    [<ffffffff8145ed16>] scsi_scan_channel+0x86/0xb0
57    [<ffffffff8145f0b0>] scsi_scan_host_selected+0x140/0x150
58    [<ffffffff8145f149>] do_scsi_scan_host+0x89/0x90
59    [<ffffffff8145f170>] do_scan_async+0x20/0x160
60    [<ffffffff8145f150>] ? do_scsi_scan_host+0x90/0x90
61    [<ffffffff810975b6>] kthread+0xa6/0xb0
62    [<ffffffff817db154>] kernel_thread_helper+0x4/0x10
63    [<ffffffff81066430>] ? finish_task_switch+0x80/0x110
64    [<ffffffff817d9c04>] ? retint_restore_args+0xe/0xe
65    [<ffffffff81097510>] ? __kthread_init_worker+0x70/0x70
66    [<ffffffff817db150>] ? gs_change+0xb/0xb
67
68Line 2776 of block/cfq-iosched.c in v3.0-rc5 is as follows::
69
70	if (rcu_dereference(ioc->ioc_data) == cic) {
71
72This form says that it must be in a plain vanilla RCU read-side critical
73section, but the "other info" list above shows that this is not the
74case.  Instead, we hold three locks, one of which might be RCU related.
75And maybe that lock really does protect this reference.  If so, the fix
76is to inform RCU, perhaps by changing __cfq_exit_single_io_context() to
77take the struct request_queue "q" from cfq_exit_queue() as an argument,
78which would permit us to invoke rcu_dereference_protected as follows::
79
80	if (rcu_dereference_protected(ioc->ioc_data,
81				      lockdep_is_held(&q->queue_lock)) == cic) {
82
83With this change, there would be no lockdep-RCU splat emitted if this
84code was invoked either from within an RCU read-side critical section
85or with the ->queue_lock held.  In particular, this would have suppressed
86the above lockdep-RCU splat because ->queue_lock is held (see #2 in the
87list above).
88
89On the other hand, perhaps we really do need an RCU read-side critical
90section.  In this case, the critical section must span the use of the
91return value from rcu_dereference(), or at least until there is some
92reference count incremented or some such.  One way to handle this is to
93add rcu_read_lock() and rcu_read_unlock() as follows::
94
95	rcu_read_lock();
96	if (rcu_dereference(ioc->ioc_data) == cic) {
97		spin_lock(&ioc->lock);
98		rcu_assign_pointer(ioc->ioc_data, NULL);
99		spin_unlock(&ioc->lock);
100	}
101	rcu_read_unlock();
102
103With this change, the rcu_dereference() is always within an RCU
104read-side critical section, which again would have suppressed the
105above lockdep-RCU splat.
106
107But in this particular case, we don't actually dereference the pointer
108returned from rcu_dereference().  Instead, that pointer is just compared
109to the cic pointer, which means that the rcu_dereference() can be replaced
110by rcu_access_pointer() as follows::
111
112	if (rcu_access_pointer(ioc->ioc_data) == cic) {
113
114Because it is legal to invoke rcu_access_pointer() without protection,
115this change would also suppress the above lockdep-RCU splat.
116