Re: rcu_preempt caused oom

["Paul E. McKenney" <paulmck@linux.ibm.com>]

On Mon, Dec 17, 2018 at 03:15:42AM +0000, He, Bo wrote:
> for double confirm the issue is not reproduce after 90 hours, we tried only add the enclosed patch on the easy reproduced build, the issue is not reproduced after 63 hours in the whole weekend on 16 boards.
> so current conclusion is the debug patch has extreme  effect on the rcu issue.

This is not a surprise.  (Please see the end of this email for a
replacement patch that won't suppress the bug.)

To see why this is not a surprise, let's take a closer look at your patch,
in light of the comment header for wait_event_idle_timeout_exclusive():

 * Returns:
 * 0 if the @condition evaluated to %false after the @timeout elapsed,
 * 1 if the @condition evaluated to %true after the @timeout elapsed,
 * or the remaining jiffies (at least 1) if the @condition evaluated
 * to %true before the @timeout elapsed.

The situation we are seeing is that the RCU_GP_FLAG_INIT is set, but
the rcu_preempt task does not wake up.  This would correspond to
the second case above, that is, a return value of 1.  Looking now
at your patch, with comments interspersed below:

------------------------------------------------------------------------

From e8b583aa685b3b4f304f72398a80461bba09389c Mon Sep 17 00:00:00 2001
From: "he, bo" <bo.he@intel.com>
Date: Sun, 9 Dec 2018 18:11:33 +0800
Subject: [PATCH] rcu: detect the preempt_rcu hang for triage jing's board

Change-Id: I2ffceec2ae4847867753609e45c99afc66956003
Tracked-On:
Signed-off-by: he, bo <bo.he@intel.com>
---
 kernel/rcu/tree.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 78c0cf2..d6de363 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2192,8 +2192,13 @@ static int __noreturn rcu_gp_kthread(void *arg)
 	int ret;
 	struct rcu_state *rsp = arg;
 	struct rcu_node *rnp = rcu_get_root(rsp);
+	pid_t rcu_preempt_pid;
 
 	rcu_bind_gp_kthread();
+	if(!strcmp(rsp->name, "rcu_preempt")) {
+		rcu_preempt_pid = rsp->gp_kthread->pid;
+	}
+
 	for (;;) {
 
 		/* Handle grace-period start. */
@@ -2202,8 +2207,19 @@ static int __noreturn rcu_gp_kthread(void *arg)
 					       READ_ONCE(rsp->gp_seq),
 					       TPS("reqwait"));
 			rsp->gp_state = RCU_GP_WAIT_GPS;
-			swait_event_idle_exclusive(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
-						     RCU_GP_FLAG_INIT);
+			if (current->pid != rcu_preempt_pid) {
+				swait_event_idle_exclusive(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
+						RCU_GP_FLAG_INIT);
+			} else {
+				ret = swait_event_idle_timeout_exclusive(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
+						RCU_GP_FLAG_INIT, 2*HZ);
+
+				if(!ret) {

We get here if ret==0.  Therefore, the above "if" statement needs to
instead be "if (ret == 1) {".

In addition, in order to get event traces dumped, we also need:

					rcu_ftrace_dump(DUMP_ALL);

+					show_rcu_gp_kthreads();
+					panic("hung_task: blocked in rcu_gp_kthread init");
+				}
+			}
+
 			rsp->gp_state = RCU_GP_DONE_GPS;
 			/* Locking provides needed memory barrier. */
 			if (rcu_gp_init(rsp))
-- 
2.7.4

------------------------------------------------------------------------

So, again, please change the "if(!ret) {" to "if (ret == 1) {", and
please add "rcu_ftrace_dump(DUMP_ALL);" right after this "if" statement,
as shown above.

With that change, I bet that you will again see failures.

> Compared with the swait_event_idle_timeout_exclusive will do 3 times to check the condition, while swait_event_idle_ exclusive will do 2 times check the condition.
> 
> so today I will do another experiment, only change as below:
> -			swait_event_idle_exclusive(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
> -						     RCU_GP_FLAG_INIT);
> +			ret = swait_event_idle_timeout_exclusive(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
> +					RCU_GP_FLAG_INIT, MAX_SCHEDULE_TIMEOUT);
> +
> 
> Can you get some clues from the experiment?

Again, please instead make the changes that I called out above, with
the replacement for your patch 0001 shown below.

							Thanx, Paul

PS.  I have been testing for quite some time, but am still unable
     to reproduce this.  So we must depend on you to reproduce it.

------------------------------------------------------------------------

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 0b760c1369f7..86152af1a580 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -2153,8 +2153,13 @@ static int __noreturn rcu_gp_kthread(void *arg)
 	int ret;
 	struct rcu_state *rsp = arg;
 	struct rcu_node *rnp = rcu_get_root(rsp);
+	pid_t rcu_preempt_pid;
 
 	rcu_bind_gp_kthread();
+	if(!strcmp(rsp->name, "rcu_preempt")) {
+		rcu_preempt_pid = rsp->gp_kthread->pid;
+	}
+
 	for (;;) {
 
 		/* Handle grace-period start. */
@@ -2163,8 +2168,20 @@ static int __noreturn rcu_gp_kthread(void *arg)
 					       READ_ONCE(rsp->gp_seq),
 					       TPS("reqwait"));
 			rsp->gp_state = RCU_GP_WAIT_GPS;
-			swait_event_idle_exclusive(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
-						     RCU_GP_FLAG_INIT);
+			if (current->pid != rcu_preempt_pid) {
+				swait_event_idle_exclusive(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
+						RCU_GP_FLAG_INIT);
+			} else {
+				ret = swait_event_idle_timeout_exclusive(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
+						RCU_GP_FLAG_INIT, 2*HZ);
+
+				if (ret == 1) {
+					rcu_ftrace_dump(DUMP_ALL);
+					show_rcu_gp_kthreads();
+					panic("hung_task: blocked in rcu_gp_kthread init");
+				}
+			}
+
 			rsp->gp_state = RCU_GP_DONE_GPS;
 			/* Locking provides needed memory barrier. */
 			if (rcu_gp_init(rsp))