Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Vaishali Thakkar]

On Friday 16 September 2016 05:40 PM, Al Viro wrote:
> On Fri, Sep 16, 2016 at 01:19:19PM +0530, Vaishali Thakkar wrote:
> 
>> Hi,
>>
>> Just a question regarding this change. As after this change
>> dput() is sleepable, is it still safe to use if under the
>> spinlock in the function d_prune_aliases?
> 
> It has always been sleepable and it wouldn't have been safe to use
> under spinlocks.  Which d_prune_aliases() does not do - __dentry_kill()
> is called with dentry, its parent and its inode (if present) all locked and
> it drops all those locks before returning.

Ah, I see. Alright. Thanks for the clarification.

> 

-- 
Vaishali

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Vaishali Thakkar]

On Friday 16 September 2016 05:40 PM, Al Viro wrote:
> On Fri, Sep 16, 2016 at 01:19:19PM +0530, Vaishali Thakkar wrote:
> 
>> Hi,
>>
>> Just a question regarding this change. As after this change
>> dput() is sleepable, is it still safe to use if under the
>> spinlock in the function d_prune_aliases?
> 
> It has always been sleepable and it wouldn't have been safe to use
> under spinlocks.  Which d_prune_aliases() does not do - __dentry_kill()
> is called with dentry, its parent and its inode (if present) all locked and
> it drops all those locks before returning.

Ah, I see. Alright. Thanks for the clarification.

> 

-- 
Vaishali

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Al Viro]

On Fri, Sep 16, 2016 at 01:19:19PM +0530, Vaishali Thakkar wrote:

> Hi,
> 
> Just a question regarding this change. As after this change
> dput() is sleepable, is it still safe to use if under the
> spinlock in the function d_prune_aliases?

It has always been sleepable and it wouldn't have been safe to use
under spinlocks.  Which d_prune_aliases() does not do - __dentry_kill()
is called with dentry, its parent and its inode (if present) all locked and
it drops all those locks before returning.

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Al Viro]

On Fri, Sep 16, 2016 at 01:19:19PM +0530, Vaishali Thakkar wrote:

> Hi,
> 
> Just a question regarding this change. As after this change
> dput() is sleepable, is it still safe to use if under the
> spinlock in the function d_prune_aliases?

It has always been sleepable and it wouldn't have been safe to use
under spinlocks.  Which d_prune_aliases() does not do - __dentry_kill()
is called with dentry, its parent and its inode (if present) all locked and
it drops all those locks before returning.

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Vaishali Thakkar]

On Wednesday 22 June 2016 08:31 AM, Wei Fang wrote:
> We triggered soft-lockup under stress test which
> open/access/write/close one file concurrently on more than
> five different CPUs:
> 
> WARN: soft lockup - CPU#0 stuck for 11s! [who:30631]
> ...
> [] dput+0x100/0x298
> [] terminate_walk+0x4c/0x60
> [] path_lookupat+0x5cc/0x7a8
> [] filename_lookup+0x38/0xf0
> [] user_path_at_empty+0x78/0xd0
> [] user_path_at+0x1c/0x28
> [] SyS_faccessat+0xb4/0x230
> 
> ->d_lock trylock may failed many times because of concurrently
> operations, and dput() may execute a long time.
> 
> Fix this by replacing cpu_relax() with cond_resched().
> dput() used to be sleepable, so make it sleepable again
> should be safe.

Hi,

Just a question regarding this change. As after this change
dput() is sleepable, is it still safe to use if under the
spinlock in the function d_prune_aliases?

Thanks

> Cc: 
> Signed-off-by: Wei Fang 
> ---
> Changes v1->v2:
> - add might_sleep() to annotate that dput() can sleep
> 
>  fs/dcache.c |4 +++-
>  1 files changed, 3 insertions(+), 1 deletions(-)
> 
> diff --git a/fs/dcache.c b/fs/dcache.c
> index d5ecc6e..074fc1c 100644
> --- a/fs/dcache.c
> +++ b/fs/dcache.c
> @@ -578,7 +578,7 @@ static struct dentry *dentry_kill(struct dentry *dentry)
>  
>  failed:
>   spin_unlock(>d_lock);
> - cpu_relax();
> + cond_resched();
>   return dentry; /* try again with same dentry */
>  }
>  
> @@ -752,6 +752,8 @@ void dput(struct dentry *dentry)
>   return;
>  
>  repeat:
> + might_sleep();
> +
>   rcu_read_lock();
>   if (likely(fast_dput(dentry))) {
>   rcu_read_unlock();
> 

-- 
Vaishali

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Vaishali Thakkar]

On Wednesday 22 June 2016 08:31 AM, Wei Fang wrote:
> We triggered soft-lockup under stress test which
> open/access/write/close one file concurrently on more than
> five different CPUs:
> 
> WARN: soft lockup - CPU#0 stuck for 11s! [who:30631]
> ...
> [] dput+0x100/0x298
> [] terminate_walk+0x4c/0x60
> [] path_lookupat+0x5cc/0x7a8
> [] filename_lookup+0x38/0xf0
> [] user_path_at_empty+0x78/0xd0
> [] user_path_at+0x1c/0x28
> [] SyS_faccessat+0xb4/0x230
> 
> ->d_lock trylock may failed many times because of concurrently
> operations, and dput() may execute a long time.
> 
> Fix this by replacing cpu_relax() with cond_resched().
> dput() used to be sleepable, so make it sleepable again
> should be safe.

Hi,

Just a question regarding this change. As after this change
dput() is sleepable, is it still safe to use if under the
spinlock in the function d_prune_aliases?

Thanks

> Cc: 
> Signed-off-by: Wei Fang 
> ---
> Changes v1->v2:
> - add might_sleep() to annotate that dput() can sleep
> 
>  fs/dcache.c |4 +++-
>  1 files changed, 3 insertions(+), 1 deletions(-)
> 
> diff --git a/fs/dcache.c b/fs/dcache.c
> index d5ecc6e..074fc1c 100644
> --- a/fs/dcache.c
> +++ b/fs/dcache.c
> @@ -578,7 +578,7 @@ static struct dentry *dentry_kill(struct dentry *dentry)
>  
>  failed:
>   spin_unlock(>d_lock);
> - cpu_relax();
> + cond_resched();
>   return dentry; /* try again with same dentry */
>  }
>  
> @@ -752,6 +752,8 @@ void dput(struct dentry *dentry)
>   return;
>  
>  repeat:
> + might_sleep();
> +
>   rcu_read_lock();
>   if (likely(fast_dput(dentry))) {
>   rcu_read_unlock();
> 

-- 
Vaishali

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Wei Fang]

Hi, Boqun,

>> diff --git a/fs/dcache.c b/fs/dcache.c
>> index d5ecc6e..074fc1c 100644
>> --- a/fs/dcache.c
>> +++ b/fs/dcache.c
>> @@ -578,7 +578,7 @@ static struct dentry *dentry_kill(struct dentry *dentry)
>>  
>>  failed:
>>  spin_unlock(>d_lock);
>> -cpu_relax();
>> +cond_resched();
> 
> Is it better to put the cond_resched() in the caller(i.e. dput()), right
> before "goto repeat"? Because it's obviously a loop there, which makes
> the purpose of cond_resched() more straightforward.

Agreed, that's more reasonable. I'll send v3 soon.

Thanks,
Wei

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Wei Fang]

Hi, Boqun,

>> diff --git a/fs/dcache.c b/fs/dcache.c
>> index d5ecc6e..074fc1c 100644
>> --- a/fs/dcache.c
>> +++ b/fs/dcache.c
>> @@ -578,7 +578,7 @@ static struct dentry *dentry_kill(struct dentry *dentry)
>>  
>>  failed:
>>  spin_unlock(>d_lock);
>> -cpu_relax();
>> +cond_resched();
> 
> Is it better to put the cond_resched() in the caller(i.e. dput()), right
> before "goto repeat"? Because it's obviously a loop there, which makes
> the purpose of cond_resched() more straightforward.

Agreed, that's more reasonable. I'll send v3 soon.

Thanks,
Wei

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Boqun Feng]

Hi Wei Fang,

On Wed, Jun 22, 2016 at 11:01:15AM +0800, Wei Fang wrote:
> We triggered soft-lockup under stress test which
> open/access/write/close one file concurrently on more than
> five different CPUs:
> 
> WARN: soft lockup - CPU#0 stuck for 11s! [who:30631]
> ...
> [] dput+0x100/0x298
> [] terminate_walk+0x4c/0x60
> [] path_lookupat+0x5cc/0x7a8
> [] filename_lookup+0x38/0xf0
> [] user_path_at_empty+0x78/0xd0
> [] user_path_at+0x1c/0x28
> [] SyS_faccessat+0xb4/0x230
> 
> ->d_lock trylock may failed many times because of concurrently
> operations, and dput() may execute a long time.
> 
> Fix this by replacing cpu_relax() with cond_resched().
> dput() used to be sleepable, so make it sleepable again
> should be safe.
> 
> Cc: 
> Signed-off-by: Wei Fang 
> ---
> Changes v1->v2:
> - add might_sleep() to annotate that dput() can sleep
> 
>  fs/dcache.c |4 +++-
>  1 files changed, 3 insertions(+), 1 deletions(-)
> 
> diff --git a/fs/dcache.c b/fs/dcache.c
> index d5ecc6e..074fc1c 100644
> --- a/fs/dcache.c
> +++ b/fs/dcache.c
> @@ -578,7 +578,7 @@ static struct dentry *dentry_kill(struct dentry *dentry)
>  
>  failed:
>   spin_unlock(>d_lock);
> - cpu_relax();
> + cond_resched();

Is it better to put the cond_resched() in the caller(i.e. dput()), right
before "goto repeat"? Because it's obviously a loop there, which makes
the purpose of cond_resched() more straightforward.

Regards,
Boqun

>   return dentry; /* try again with same dentry */
>  }
>  
> @@ -752,6 +752,8 @@ void dput(struct dentry *dentry)
>   return;
>  
>  repeat:
> + might_sleep();
> +
>   rcu_read_lock();
>   if (likely(fast_dput(dentry))) {
>   rcu_read_unlock();
> -- 
> 1.7.1
> 


signature.asc
Description: PGP signature

Re: [PATCH v2] fs/dcache.c: avoid soft-lockup in dput()

[Boqun Feng]

Hi Wei Fang,

On Wed, Jun 22, 2016 at 11:01:15AM +0800, Wei Fang wrote:
> We triggered soft-lockup under stress test which
> open/access/write/close one file concurrently on more than
> five different CPUs:
> 
> WARN: soft lockup - CPU#0 stuck for 11s! [who:30631]
> ...
> [] dput+0x100/0x298
> [] terminate_walk+0x4c/0x60
> [] path_lookupat+0x5cc/0x7a8
> [] filename_lookup+0x38/0xf0
> [] user_path_at_empty+0x78/0xd0
> [] user_path_at+0x1c/0x28
> [] SyS_faccessat+0xb4/0x230
> 
> ->d_lock trylock may failed many times because of concurrently
> operations, and dput() may execute a long time.
> 
> Fix this by replacing cpu_relax() with cond_resched().
> dput() used to be sleepable, so make it sleepable again
> should be safe.
> 
> Cc: 
> Signed-off-by: Wei Fang 
> ---
> Changes v1->v2:
> - add might_sleep() to annotate that dput() can sleep
> 
>  fs/dcache.c |4 +++-
>  1 files changed, 3 insertions(+), 1 deletions(-)
> 
> diff --git a/fs/dcache.c b/fs/dcache.c
> index d5ecc6e..074fc1c 100644
> --- a/fs/dcache.c
> +++ b/fs/dcache.c
> @@ -578,7 +578,7 @@ static struct dentry *dentry_kill(struct dentry *dentry)
>  
>  failed:
>   spin_unlock(>d_lock);
> - cpu_relax();
> + cond_resched();

Is it better to put the cond_resched() in the caller(i.e. dput()), right
before "goto repeat"? Because it's obviously a loop there, which makes
the purpose of cond_resched() more straightforward.

Regards,
Boqun

>   return dentry; /* try again with same dentry */
>  }
>  
> @@ -752,6 +752,8 @@ void dput(struct dentry *dentry)
>   return;
>  
>  repeat:
> + might_sleep();
> +
>   rcu_read_lock();
>   if (likely(fast_dput(dentry))) {
>   rcu_read_unlock();
> -- 
> 1.7.1
> 


signature.asc
Description: PGP signature