ANDROID: nf: IDLETIMER: Fix possible use before initialization in idletimer_resume

idletimer_resume() assumes that the PM_SUSPEND_PREPARE notifier is sent
before PM_POST_SUSPEND so that timer->last_suspend_time is initialized.
However, it is possible for PM_POST_SUSPEND to be sent first if there is an
error returned from another driver's PM_SUSPEND_PREPARE notifier. As a
result, structures are accessed without initialization.

The arguments to set_normalized_timespec are very large and unexpected.
timer->last_suspend_time has the value of
{.tv_sec = 0x6b6b6b6b6b6b6b6b, .tv_nsec=0x6b6b6b6b6b6b6b6b}. Since really
large iterations are required, this operation takes more than a minute
and causes the CPU to trigger a spinbug since the timestamp lock is held.

Call stack -

- set_normalized_timespec
- timespec_sub
- idletimer_resume
- notifier_call_chain
- __blocking_notifier_call_chain
- pm_notifier_call_chain

Add a flag indicating whether the current value of timer->last_suspend is
valid.

Detected with CONFIG_SLUB_DEBUG & CONFIG_DEBUG_SPINLOCK in arm64.

Bug: 140404598
Fixes: f0c2df2b1228a ("ANDROID: netfilter: xt_IDLETIMER: Add new
netlink msg type")
Change-Id: I95328b0ac85dba819ff9cef751c3d07300c232f1
Signed-off-by: Subash Abhinov Kasiviswanathan <subashab@codeaurora.org>
Signed-off-by: Todd Kjos <tkjos@google.com>
(cherry picked from commit 7ded4359680d3fb593cbc5c90d84bb5e5083eda9)
This commit is contained in:
Subash Abhinov Kasiviswanathan 2016-11-10 19:18:53 -07:00 committed by Maciej Żenczykowski
parent 96267fbb74
commit 3acdbda6a7

View file

@ -70,6 +70,7 @@ struct idletimer_tg {
bool send_nl_msg;
bool active;
uid_t uid;
bool suspend_time_valid;
};
static LIST_HEAD(idletimer_tg_list);
@ -239,8 +240,13 @@ static int idletimer_resume(struct notifier_block *notifier,
switch (pm_event) {
case PM_SUSPEND_PREPARE:
get_monotonic_boottime(&timer->last_suspend_time);
timer->suspend_time_valid = true;
break;
case PM_POST_SUSPEND:
if (!timer->suspend_time_valid)
break;
timer->suspend_time_valid = false;
spin_lock_bh(&timestamp_lock);
if (!timer->active) {
spin_unlock_bh(&timestamp_lock);
@ -291,7 +297,7 @@ static int idletimer_tg_create(struct idletimer_tg_info *info)
{
int ret;
info->timer = kmalloc(sizeof(*info->timer), GFP_KERNEL);
info->timer = kzalloc(sizeof(*info->timer), GFP_KERNEL);
if (!info->timer) {
ret = -ENOMEM;
goto out;