From: Rocco Yue <[email protected]>
We often encounter system hangs caused by certain process
holding rtnl_lock for a long time. Even if there is a lock
detection mechanism in Linux, it is a bit troublesome and
affects the system performance. We hope to add a lightweight
debugging mechanism for detecting rtnl_lock.
Up to now, we have discovered and solved some potential bugs
through this lightweight rtnl_lock debugging mechanism, which
is helpful for us.
When you say Y for RTNL_LOCK_DEBUG, then the kernel will
detect if any function hold rtnl_lock too long and some key
information will be printed out to help locate the problem.
i.e: from the following logs, we can clearly know that the
pid=2206 RfxSender_4 process holds rtnl_lock for a long time,
causing the system to hang. And we can also speculate that the
delay operation may be performed in devinet_ioctl(), resulting
in rtnl_lock was not released in time.
[ 40.191481] rtnetlink: -- rtnl_print_btrace start --
[ 40.191494] rtnetlink: RfxSender_4[2206][R] hold rtnl_lock
more than 2 sec, start time: 38181400013
[ 40.191571] Call trace:
[ 40.191586] rtnl_print_btrace+0xf0/0x124
[ 40.191656] __delay+0xc0/0x180
[ 40.191663] devinet_ioctl+0x21c/0x75c
[ 40.191668] inet_ioctl+0xb8/0x1f8
[ 40.191675] sock_do_ioctl+0x70/0x2ac
[ 40.191682] sock_ioctl+0x5dc/0xa74
[ 40.191715] rtnetlink: -- rtnl_print_btrace end --
[ 42.181879] rtnetlink: rtnl_lock is held by [2206] from
[38181400013] to [42181875177]
Signed-off-by: Rocco Yue <[email protected]>
---
lib/Kconfig.debug | 10 ++++++
net/core/rtnetlink.c | 85 ++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 95 insertions(+)
diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
index 678c13967580..f1a722e16bee 100644
--- a/lib/Kconfig.debug
+++ b/lib/Kconfig.debug
@@ -2027,6 +2027,16 @@ config KCOV_IRQ_AREA_SIZE
soft interrupts. This specifies the size of those areas in the
number of unsigned long words.
+config RTNL_LOCK_DEBUG
+ bool "rtnl_lock debugging, deadlock detection"
+ depends on STACKTRACE_SUPPORT
+ select STACKTRACE
+ help
+ If you say Y here then the kernel will detect whether any function
+ hold rtnl_lock too long and some key information will be printed
+ out to help locate the problem.
+ If unsure, say N.
+
menuconfig RUNTIME_TESTING_MENU
bool "Runtime Testing"
def_bool y
diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index 714d5fa38546..4f81086e5a42 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c
@@ -12,6 +12,8 @@
* Vitaly E. Lavrov RTA_OK arithmetics was wrong.
*/
+#define pr_fmt(fmt) "rtnetlink: " fmt
+
#include <linux/bitops.h>
#include <linux/errno.h>
#include <linux/module.h>
@@ -57,6 +59,81 @@
#define RTNL_MAX_TYPE 50
#define RTNL_SLAVE_MAX_TYPE 40
+#ifdef CONFIG_RTNL_LOCK_DEBUG
+
+#include <linux/sched/debug.h>
+#include <linux/stacktrace.h>
+
+/* Debug log and btrace will be printed when the rtnl_lock
+ * is held for more than RTNL_LOCK_MAX_HOLD_TIME seconds
+ */
+#define RTNL_LOCK_MAX_HOLD_TIME 2
+
+#define RTNL_LOCK_MAX_TRACE 10 /* stack trace length */
+
+struct rtnl_debug_btrace_t {
+ struct task_struct *task;
+ int pid;
+ unsigned long long start_time;
+ unsigned long long end_time;
+ unsigned long addrs[RTNL_LOCK_MAX_TRACE];
+ unsigned int nr_entries;
+};
+
+static struct rtnl_debug_btrace_t rtnl_instance;
+
+static void rtnl_print_btrace(struct timer_list *unused);
+static DEFINE_TIMER(rtnl_chk_timer, rtnl_print_btrace);
+
+/* Save stack trace to the given array of RTNL_LOCK_MAX_TRACE size.
+ */
+static int __save_stack_trace(unsigned long *trace)
+{
+ return stack_trace_save(trace, RTNL_LOCK_MAX_TRACE, 0);
+}
+
+static void rtnl_get_btrace(struct task_struct *who)
+{
+ unsigned long expires;
+
+ rtnl_instance.task = who;
+ rtnl_instance.pid = who->pid;
+ rtnl_instance.start_time = sched_clock();
+ rtnl_instance.end_time = 0;
+ rtnl_instance.nr_entries = __save_stack_trace(rtnl_instance.addrs);
+
+ expires = jiffies + RTNL_LOCK_MAX_HOLD_TIME * HZ;
+ mod_timer(&rtnl_chk_timer, expires);
+}
+
+static void rtnl_print_btrace(struct timer_list *unused)
+{
+ pr_info("-- %s start --\n", __func__);
+ pr_info("%s[%d][%c] hold rtnl_lock more than %d sec, start time: %llu\n",
+ rtnl_instance.task->comm,
+ rtnl_instance.pid,
+ task_state_to_char(rtnl_instance.task),
+ RTNL_LOCK_MAX_HOLD_TIME,
+ rtnl_instance.start_time);
+ stack_trace_print(rtnl_instance.addrs, rtnl_instance.nr_entries, 0);
+ show_stack(rtnl_instance.task, NULL, KERN_INFO);
+ pr_info("-- %s end --\n", __func__);
+}
+
+static void rtnl_relase_btrace(void)
+{
+ rtnl_instance.end_time = sched_clock();
+ del_timer_sync(&rtnl_chk_timer);
+
+ if (rtnl_instance.end_time - rtnl_instance.start_time > 2 * NSEC_PER_SEC) {
+ pr_info("rtnl_lock is held by [%d] from [%llu] to [%llu]\n",
+ rtnl_instance.pid,
+ rtnl_instance.start_time,
+ rtnl_instance.end_time);
+ }
+}
+#endif
+
struct rtnl_link {
rtnl_doit_func doit;
rtnl_dumpit_func dumpit;
@@ -70,6 +147,10 @@ static DEFINE_MUTEX(rtnl_mutex);
void rtnl_lock(void)
{
mutex_lock(&rtnl_mutex);
+
+#ifdef CONFIG_RTNL_LOCK_DEBUG
+ rtnl_get_btrace(current);
+#endif
}
EXPORT_SYMBOL(rtnl_lock);
@@ -95,6 +176,10 @@ void __rtnl_unlock(void)
defer_kfree_skb_list = NULL;
+#ifdef CONFIG_RTNL_LOCK_DEBUG
+ rtnl_relase_btrace();
+#endif
+
mutex_unlock(&rtnl_mutex);
while (head) {
--
2.18.0
On Tue, May 11, 2021 at 4:46 AM Rocco yue <[email protected]> wrote:
>
> From: Rocco Yue <[email protected]>
>
> We often encounter system hangs caused by certain process
> holding rtnl_lock for a long time. Even if there is a lock
> detection mechanism in Linux, it is a bit troublesome and
> affects the system performance. We hope to add a lightweight
> debugging mechanism for detecting rtnl_lock.
Any reason why this is specific to RTNL lock? To me holding
a mutex lock for a long time is problematic for any mutex.
I have seen some fs mutex being held for a long time caused
many hung tasks in the system.
Thanks.
On Tue, 11 May 2021 19:32:57 +0800 Rocco yue wrote:
> We often encounter system hangs caused by certain process
> holding rtnl_lock for a long time. Even if there is a lock
> detection mechanism in Linux, it is a bit troublesome and
> affects the system performance. We hope to add a lightweight
> debugging mechanism for detecting rtnl_lock.
>
> Up to now, we have discovered and solved some potential bugs
> through this lightweight rtnl_lock debugging mechanism, which
> is helpful for us.
>
> When you say Y for RTNL_LOCK_DEBUG, then the kernel will
> detect if any function hold rtnl_lock too long and some key
> information will be printed out to help locate the problem.
>
> i.e: from the following logs, we can clearly know that the
> pid=2206 RfxSender_4 process holds rtnl_lock for a long time,
> causing the system to hang. And we can also speculate that the
> delay operation may be performed in devinet_ioctl(), resulting
> in rtnl_lock was not released in time.
You can achieve that with a pair of fexit/fentry hooks or kprobes,
and maybe a bit of BPF. No need for config options, and hardcoded
parameters..
On Wed, May 12, 2021 at 1:38 AM Rocco.Yue <[email protected]> wrote:
>
> On Tue, 2021-05-11 at 10:00 -0700, Cong Wang wrote:
> > On Tue, May 11, 2021 at 4:46 AM Rocco yue <[email protected]> wrote:
> > >
> > > From: Rocco Yue <[email protected]>
> > >
> > > We often encounter system hangs caused by certain process
> > > holding rtnl_lock for a long time. Even if there is a lock
> > > detection mechanism in Linux, it is a bit troublesome and
> > > affects the system performance. We hope to add a lightweight
> > > debugging mechanism for detecting rtnl_lock.
> >
> > Any reason why this is specific to RTNL lock? To me holding
> > a mutex lock for a long time is problematic for any mutex.
> > I have seen some fs mutex being held for a long time caused
> > many hung tasks in the system.
> >
> > Thanks.
>
> Thank you for a good question.
>
> It's a problem to hold rtnl_lock for a long time, and other locks are no
> exception.
Then please try to make it generic so that other mutex locks will
benefit too.
Thanks.