diff mbox series

mm/mmu_notifiers: use helper function mmu_notifier_synchronize()

Message ID 20220217110948.35477-1-linmiaohe@huawei.com (mailing list archive)
State New
Headers show
Series mm/mmu_notifiers: use helper function mmu_notifier_synchronize() | expand

Commit Message

Miaohe Lin Feb. 17, 2022, 11:09 a.m. UTC
Use helper function mmu_notifier_synchronize() to ensure all mmu_notifiers
are freed. Minor readability improvement.

Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
---
 mm/mmu_notifier.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

Comments

Matthew Wilcox Feb. 17, 2022, 1:32 p.m. UTC | #1
On Thu, Feb 17, 2022 at 07:09:48PM +0800, Miaohe Lin wrote:
> Use helper function mmu_notifier_synchronize() to ensure all mmu_notifiers
> are freed. Minor readability improvement.

Is it though?

> @@ -334,15 +334,15 @@ static void mn_hlist_release(struct mmu_notifier_subscriptions *subscriptions,
>  	srcu_read_unlock(&srcu, id);
>  
>  	/*
> -	 * synchronize_srcu here prevents mmu_notifier_release from returning to
> -	 * exit_mmap (which would proceed with freeing all pages in the mm)
> -	 * until the ->release method returns, if it was invoked by
> -	 * mmu_notifier_unregister.
> +	 * mmu_notifier_synchronize here prevents mmu_notifier_release from
> +	 * returning to exit_mmap (which would proceed with freeing all pages
> +	 * in the mm) until the ->release method returns, if it was invoked
> +	 * by mmu_notifier_unregister.
>  	 *
>  	 * The notifier_subscriptions can't go away from under us because
>  	 * one mm_count is held by exit_mmap.
>  	 */
> -	synchronize_srcu(&srcu);
> +	mmu_notifier_synchronize();

We just read_unlocked the &srcu.  Now I have to jump to the definition
of mmu_notifier_synchronize() to find out that it's now waiting for the
very same srcu.  I think this abstraction makes the code harder to read,
not easier.

>  }
>  
>  void __mmu_notifier_release(struct mm_struct *mm)
> @@ -851,7 +851,7 @@ void mmu_notifier_unregister(struct mmu_notifier *subscription,
>  	 * Wait for any running method to finish, of course including
>  	 * ->release if it was run by mmu_notifier_release instead of us.
>  	 */
> -	synchronize_srcu(&srcu);
> +	mmu_notifier_synchronize();

Same here.
Miaohe Lin Feb. 17, 2022, 1:44 p.m. UTC | #2
On 2022/2/17 21:32, Matthew Wilcox wrote:
> On Thu, Feb 17, 2022 at 07:09:48PM +0800, Miaohe Lin wrote:
>> Use helper function mmu_notifier_synchronize() to ensure all mmu_notifiers
>> are freed. Minor readability improvement.
> 
> Is it though?
> 
>> @@ -334,15 +334,15 @@ static void mn_hlist_release(struct mmu_notifier_subscriptions *subscriptions,
>>  	srcu_read_unlock(&srcu, id);
>>  
>>  	/*
>> -	 * synchronize_srcu here prevents mmu_notifier_release from returning to
>> -	 * exit_mmap (which would proceed with freeing all pages in the mm)
>> -	 * until the ->release method returns, if it was invoked by
>> -	 * mmu_notifier_unregister.
>> +	 * mmu_notifier_synchronize here prevents mmu_notifier_release from
>> +	 * returning to exit_mmap (which would proceed with freeing all pages
>> +	 * in the mm) until the ->release method returns, if it was invoked
>> +	 * by mmu_notifier_unregister.
>>  	 *
>>  	 * The notifier_subscriptions can't go away from under us because
>>  	 * one mm_count is held by exit_mmap.
>>  	 */
>> -	synchronize_srcu(&srcu);
>> +	mmu_notifier_synchronize();
> 
> We just read_unlocked the &srcu.  Now I have to jump to the definition
> of mmu_notifier_synchronize() to find out that it's now waiting for the
> very same srcu.  I think this abstraction makes the code harder to read,
> not easier.
> 

From this point of view, this helper would disturb the understanding of the code.
Many thanks for pointing this out. Sorry for my mindlessness.

>>  }
>>  
>>  void __mmu_notifier_release(struct mm_struct *mm)
>> @@ -851,7 +851,7 @@ void mmu_notifier_unregister(struct mmu_notifier *subscription,
>>  	 * Wait for any running method to finish, of course including
>>  	 * ->release if it was run by mmu_notifier_release instead of us.
>>  	 */
>> -	synchronize_srcu(&srcu);
>> +	mmu_notifier_synchronize();
> 
> Same here.
> 
> .
>
Jason Gunthorpe Feb. 17, 2022, 3:50 p.m. UTC | #3
On Thu, Feb 17, 2022 at 07:09:48PM +0800, Miaohe Lin wrote:
> Use helper function mmu_notifier_synchronize() to ensure all mmu_notifiers
> are freed. Minor readability improvement.
> 
> Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
> ---
>  mm/mmu_notifier.c | 12 ++++++------
>  1 file changed, 6 insertions(+), 6 deletions(-)

I'm not keen on this, the internal synchronize_srcu's don't have the
same usage model as described in the comment for
mmu_notifier_synchronize(). Instead they are doing what their comments
say.

Yes, it is the same code, but the purpose is different.

Jason
Miaohe Lin Feb. 18, 2022, 1:49 a.m. UTC | #4
On 2022/2/17 23:50, Jason Gunthorpe wrote:
> On Thu, Feb 17, 2022 at 07:09:48PM +0800, Miaohe Lin wrote:
>> Use helper function mmu_notifier_synchronize() to ensure all mmu_notifiers
>> are freed. Minor readability improvement.
>>
>> Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
>> ---
>>  mm/mmu_notifier.c | 12 ++++++------
>>  1 file changed, 6 insertions(+), 6 deletions(-)
> 
> I'm not keen on this, the internal synchronize_srcu's don't have the
> same usage model as described in the comment for
> mmu_notifier_synchronize(). Instead they are doing what their comments
> say.
> 
> Yes, it is the same code, but the purpose is different.
> 

Yep, this is my overlook. Many thanks for reply.

> Jason
> .
>
diff mbox series

Patch

diff --git a/mm/mmu_notifier.c b/mm/mmu_notifier.c
index 459d195d2ff6..159f70c20236 100644
--- a/mm/mmu_notifier.c
+++ b/mm/mmu_notifier.c
@@ -334,15 +334,15 @@  static void mn_hlist_release(struct mmu_notifier_subscriptions *subscriptions,
 	srcu_read_unlock(&srcu, id);
 
 	/*
-	 * synchronize_srcu here prevents mmu_notifier_release from returning to
-	 * exit_mmap (which would proceed with freeing all pages in the mm)
-	 * until the ->release method returns, if it was invoked by
-	 * mmu_notifier_unregister.
+	 * mmu_notifier_synchronize here prevents mmu_notifier_release from
+	 * returning to exit_mmap (which would proceed with freeing all pages
+	 * in the mm) until the ->release method returns, if it was invoked
+	 * by mmu_notifier_unregister.
 	 *
 	 * The notifier_subscriptions can't go away from under us because
 	 * one mm_count is held by exit_mmap.
 	 */
-	synchronize_srcu(&srcu);
+	mmu_notifier_synchronize();
 }
 
 void __mmu_notifier_release(struct mm_struct *mm)
@@ -851,7 +851,7 @@  void mmu_notifier_unregister(struct mmu_notifier *subscription,
 	 * Wait for any running method to finish, of course including
 	 * ->release if it was run by mmu_notifier_release instead of us.
 	 */
-	synchronize_srcu(&srcu);
+	mmu_notifier_synchronize();
 
 	BUG_ON(atomic_read(&mm->mm_count) <= 0);