[v3] rtnetlink: add rtnl_lock debug log

Commit Message

Rocco Yue May 11, 2021, 11:32 a.m. UTC

From: Rocco Yue <[e-mail]>

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 <[e-mail]>
---
 lib/Kconfig.debug    | 10 ++++++
 net/core/rtnetlink.c | 85 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 95 insertions(+)

Comments

Cong Wang May 11, 2021, 5 p.m. UTC | #1

On Tue, May 11, 2021 at 4:46 AM Rocco yue <[e-mail]> wrote:
>
> From: Rocco Yue <[e-mail]>
>
> 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.

Jakub Kicinski May 12, 2021, 3:39 p.m. UTC | #2

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..

Cong Wang May 12, 2021, 4:15 p.m. UTC | #3

On Wed, May 12, 2021 at 1:38 AM Rocco.Yue <[e-mail]> wrote:
>
> On Tue, 2021-05-11 at 10:00 -0700, Cong Wang wrote:
> > On Tue, May 11, 2021 at 4:46 AM Rocco yue <[e-mail]> wrote:
> > >
> > > From: Rocco Yue <[e-mail]>
> > >
> > > 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.

Patch

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) {