2023-01-11 13:25:24

by Zqiang

[permalink] [raw]
Subject: [PATCH] rcu: Fix the start_poll_synchronize_rcu_expedited() be invoked very early

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 <[email protected]>
---
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


2023-01-11 15:24:02

by Frederic Weisbecker

[permalink] [raw]
Subject: Re: [PATCH] rcu: Fix the start_poll_synchronize_rcu_expedited() be invoked very early

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 <[email protected]>
> ---
> 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.

2023-01-12 02:39:33

by Zqiang

[permalink] [raw]
Subject: RE: [PATCH] rcu: Fix the start_poll_synchronize_rcu_expedited() be invoked very early

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 <[email protected]>
> ---
> 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.