Currently, the start_poll_synchronize_rcu_expedited() can be invoked
very early. before rcu_init(), the rcu_data structure's->mynode is not
initialized, if invoke start_poll_synchronize_rcu_expedited() before
rcu_init(), will access to NULL mynode pointer.

This commit therefore add exp_seq_poll_rq member to rcu_state structure
to store snap seq number

Signed-off-by: Zqiang <qiang1.zhang@intel.com>
---
 kernel/rcu/tree.c     | 3 ++-
 kernel/rcu/tree.h     | 1 +
 kernel/rcu/tree_exp.h | 6 ++++--
 3 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 63545d79da51..34b13d6bd8c4 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -92,6 +92,7 @@ static struct rcu_state rcu_state = {
 	.exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
 	.exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
 	.ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
+	.exp_seq_poll_rq = RCU_GET_STATE_COMPLETED,
 };
 
 /* Dump rcu_node combining tree at boot to verify correct setup. */
@@ -4938,7 +4939,7 @@ void __init rcu_init(void)
 		qovld_calc = qovld;
 
 	// Kick-start any polled grace periods that started early.
-	if (!(per_cpu_ptr(&rcu_data, cpu)->mynode->exp_seq_poll_rq & 0x1))
+	if (!(rcu_state.exp_seq_poll_rq & 0x1))
 		(void)start_poll_synchronize_rcu_expedited();
 
 	rcu_test_sync_prims();
diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h
index 192536916f9a..dbc7c7484a7e 100644
--- a/kernel/rcu/tree.h
+++ b/kernel/rcu/tree.h
@@ -397,6 +397,7 @@ struct rcu_state {
 						/* Synchronize offline with */
 						/*  GP pre-initialization. */
 	int nocb_is_setup;			/* nocb is setup from boot */
+	unsigned long exp_seq_poll_rq;
 };
 
 /* Values for rcu_state structure's gp_flags field. */
diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
index 956cd459ba7f..5964d1cccab1 100644
--- a/kernel/rcu/tree_exp.h
+++ b/kernel/rcu/tree_exp.h
@@ -1068,9 +1068,11 @@ unsigned long start_poll_synchronize_rcu_expedited(void)
 	if (rcu_init_invoked())
 		raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
 	if (!poll_state_synchronize_rcu(s)) {
-		rnp->exp_seq_poll_rq = s;
-		if (rcu_init_invoked())
+		if (rcu_init_invoked()) {
+			rnp->exp_seq_poll_rq = s;
 			queue_work(rcu_gp_wq, &rnp->exp_poll_wq);
+		} else
+			rcu_state.exp_seq_poll_rq = s;
 	}
 	if (rcu_init_invoked())
 		raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
-- 
2.25.1

On Wed, Jan 11, 2023 at 09:14:53PM +0800, Zqiang wrote:
> Currently, the start_poll_synchronize_rcu_expedited() can be invoked
> very early. before rcu_init(), the rcu_data structure's->mynode is not
> initialized, if invoke start_poll_synchronize_rcu_expedited() before
> rcu_init(), will access to NULL mynode pointer.
> 
> This commit therefore add exp_seq_poll_rq member to rcu_state structure
> to store snap seq number

Is it even sane to poll that early in the morning? :-)

> 
> Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> ---
>  kernel/rcu/tree.c     | 3 ++-
>  kernel/rcu/tree.h     | 1 +
>  kernel/rcu/tree_exp.h | 6 ++++--
>  3 files changed, 7 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 63545d79da51..34b13d6bd8c4 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -92,6 +92,7 @@ static struct rcu_state rcu_state = {
>  	.exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
>  	.exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
>  	.ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
> +	.exp_seq_poll_rq = RCU_GET_STATE_COMPLETED,

I don't know if we really want to fix this, but assuming we do,
can we rename it to boot_exp_seq_poll_rq? To avoid later confusion.

Thanks.

On Wed, Jan 11, 2023 at 09:14:53PM +0800, Zqiang wrote:
> Currently, the start_poll_synchronize_rcu_expedited() can be invoked
> very early. before rcu_init(), the rcu_data structure's->mynode is not
> initialized, if invoke start_poll_synchronize_rcu_expedited() before
> rcu_init(), will access to NULL mynode pointer.
> 
> This commit therefore add exp_seq_poll_rq member to rcu_state structure
> to store snap seq number
>
>Is it even sane to poll that early in the morning? :-)

According to d96c52fe4907c ("rcu: Add polled expedited grace-period primitives "),
the start_poll_synchronize_rcu_expedited() can be invoked very early, that is to say
can be invoked before rcu_init(),  the following code snippet also shows that the 
start_poll_synchronize_rcu_expedited()  may be called early.

// Kick-start any polled grace periods that started early.
         if (!(per_cpu_ptr(&rcu_data, cpu)->mynode->exp_seq_poll_rq & 0x1))
                 (void)start_poll_synchronize_rcu_expedited();

But before rcu_init(), the rcu_data structure's->mynode is not initialized,
so at least, we set rnp->exp_seq_poll_rq should be in case the rcu_init_invoked()
returns true in start_poll_synchronize_rcu_expedited() .


>
> 
> Signed-off-by: Zqiang <qiang1.zhang@intel.com>
> ---
>  kernel/rcu/tree.c     | 3 ++-
>  kernel/rcu/tree.h     | 1 +
>  kernel/rcu/tree_exp.h | 6 ++++--
>  3 files changed, 7 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 63545d79da51..34b13d6bd8c4 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -92,6 +92,7 @@ static struct rcu_state rcu_state = {
>  	.exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
>  	.exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
>  	.ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
> +	.exp_seq_poll_rq = RCU_GET_STATE_COMPLETED,
>
>I don't know if we really want to fix this, but assuming we do,
>can we rename it to boot_exp_seq_poll_rq? To avoid later confusion.

Will fix in next version.

Thanks
Zqiang

>
>Thanks.