2021-10-08 03:28:07

by Xu, Yanfei

Subject: [PATCH v2] locking/mutex: remove rcu_read_lock/unlock as we already disabled preemption

preempt_disable/enable() is equal to RCU read-side crital section,
and the mutex lock slowpath disabled the preemption for the optimistic
spinning code. Let's remove the rcu_read_lock/unlock for saving some
cycles in hot codes.

Signed-off-by: Yanfei Xu <[email protected]>
---
v1->v2: fix the incorrect comment in code and commit message.
thanks for WaiMan's suggestion.

BTW, sorry for this late v2 due to a long vocation.

kernel/locking/mutex.c | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/kernel/locking/mutex.c b/kernel/locking/mutex.c
index 2fede72b6af5..2f654cfb10d9 100644
--- a/kernel/locking/mutex.c
+++ b/kernel/locking/mutex.c
@@ -351,13 +351,14 @@ bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
{
bool ret = true;

- rcu_read_lock();
while (__mutex_owner(lock) == owner) {
/*
* Ensure we emit the owner->on_cpu, dereference _after_
- * checking lock->owner still matches owner. If that fails,
- * owner might point to freed memory. If it still matches,
- * the rcu_read_lock() ensures the memory stays valid.
+ * checking lock->owner still matches owner. And we already
+ * disabled preemption which is equal to the RCU read-side
+ * crital section in optimistic spinning code. Thus the
+ * task_strcut structure won't go away during the spinning
+ * period
*/
barrier();

@@ -377,7 +378,6 @@ bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,

cpu_relax();
}
- rcu_read_unlock();

return ret;
}
--
2.27.0

2021-10-12 13:20:34

by Peter Zijlstra

Subject: Re: [PATCH v2] locking/mutex: remove rcu_read_lock/unlock as we already disabled preemption

On Fri, Oct 08, 2021 at 11:25:18AM +0800, Yanfei Xu wrote:
> preempt_disable/enable() is equal to RCU read-side crital section,
> and the mutex lock slowpath disabled the preemption for the optimistic
> spinning code. Let's remove the rcu_read_lock/unlock for saving some
> cycles in hot codes.
>
> Signed-off-by: Yanfei Xu <[email protected]>
> ---
> v1->v2: fix the incorrect comment in code and commit message.
> thanks for WaiMan's suggestion.
>
> BTW, sorry for this late v2 due to a long vocation.
>
> kernel/locking/mutex.c | 10 +++++-----
> 1 file changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/kernel/locking/mutex.c b/kernel/locking/mutex.c
> index 2fede72b6af5..2f654cfb10d9 100644
> --- a/kernel/locking/mutex.c
> +++ b/kernel/locking/mutex.c
> @@ -351,13 +351,14 @@ bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
> {
> bool ret = true;

lockdep_assert_preemption_disabled();

> - rcu_read_lock();
> while (__mutex_owner(lock) == owner) {
> /*
> * Ensure we emit the owner->on_cpu, dereference _after_

And did you check the other code in locking/ for similar things?

2021-10-13 07:56:29

by Xu, Yanfei

Subject: Re: [PATCH v2] locking/mutex: remove rcu_read_lock/unlock as we already disabled preemption

On 10/12/21 9:15 PM, Peter Zijlstra wrote:
> [Please note: This e-mail is from an EXTERNAL e-mail address]
>
> On Fri, Oct 08, 2021 at 11:25:18AM +0800, Yanfei Xu wrote:
>> preempt_disable/enable() is equal to RCU read-side crital section,
>> and the mutex lock slowpath disabled the preemption for the optimistic
>> spinning code. Let's remove the rcu_read_lock/unlock for saving some
>> cycles in hot codes.
>>
>> Signed-off-by: Yanfei Xu <[email protected]>
>> ---
>> v1->v2: fix the incorrect comment in code and commit message.
>> thanks for WaiMan's suggestion.
>>
>> BTW, sorry for this late v2 due to a long vocation.
>>
>> kernel/locking/mutex.c | 10 +++++-----
>> 1 file changed, 5 insertions(+), 5 deletions(-)
>>
>> diff --git a/kernel/locking/mutex.c b/kernel/locking/mutex.c
>> index 2fede72b6af5..2f654cfb10d9 100644
>> --- a/kernel/locking/mutex.c
>> +++ b/kernel/locking/mutex.c
>> @@ -351,13 +351,14 @@ bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
>> {
>> bool ret = true;
>
> lockdep_assert_preemption_disabled();

Agree.

>
>> - rcu_read_lock();
>> while (__mutex_owner(lock) == owner) {
>> /*
>> * Ensure we emit the owner->on_cpu, dereference _after_
>
> And did you check the other code in locking/ for similar things?
>

I did a check, rwsem also has the similar things. Will do it for rwsem
in v3.



Thanks,
Yanfei