[RFC][PATCH 1/2] fs: use RCU for free_super() vs. __sb_start_write()

[RFC][PATCH 1/2] fs: use RCU for free_super() vs. __sb_start_write()

[Dave Hansen]

Currently, __sb_start_write() and freeze_super() can race with
each other.  __sb_start_write() uses a smp_mb() to ensure that
freeze_super() can see its write to sb->s_writers.counter and
that it can see freeze_super()'s update to sb->s_writers.frozen.
This all seems to work fine.

But, this smp_mb() makes __sb_start_write() the single hottest
function in the kernel if I sit in a loop and do tiny write()s to
tmpfs over and over.  This is on a very small 2-core system, so
it will only get worse on larger systems.

This _seems_ like an ideal case for RCU.  __sb_start_write() is
the RCU read-side and is in a very fast, performance-sensitive
path.  freeze_super() is the RCU writer and is in an extremely
rare non-performance-sensitive path.

Instead of doing and smp_wmb() in __sb_start_write(), we do
rcu_read_lock().  This ensures that a CPU doing freeze_super()
can not proceed past its synchronize_rcu() until the grace
period has ended and the 's_writers.frozen = SB_FREEZE_WRITE'
is visible to __sb_start_write().

One question here: Does the work that __sb_start_write() does in
a previous grace period becomes visible to freeze_super() after
its call to synchronize_rcu()?  It _seems_ like it should, but it
seems backwards to me since __sb_start_write() is the "reader" in
this case.

This patch increases the number of writes/second that I can do
by 10.4%.

Does anybody see any holes with this?

Cc: Jan Kara <jack@suse.cz>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>

---

 b/fs/super.c |   38 +++++++++++++++++++++-----------------
 1 file changed, 21 insertions(+), 17 deletions(-)

diff -puN fs/super.c~rcu-__sb_start_write fs/super.c
--- a/fs/super.c~rcu-__sb_start_write	2015-06-19 14:50:53.081869092 -0700
+++ b/fs/super.c	2015-06-19 15:19:03.000473047 -0700
@@ -1190,27 +1190,25 @@ static void acquire_freeze_lock(struct s
  */
 int __sb_start_write(struct super_block *sb, int level, bool wait)
 {
-retry:
-	if (unlikely(sb->s_writers.frozen >= level)) {
+	/*
+	 * RCU keeps freeze_super() from proceeding
+	 * while we are in here.
+	 */
+	rcu_read_lock();
+	while (unlikely(sb->s_writers.frozen >= level)) {
+		rcu_read_unlock();
 		if (!wait)
-			return 0;
+	       		return 0;
 		wait_event(sb->s_writers.wait_unfrozen,
 			   sb->s_writers.frozen < level);
+		rcu_read_lock();
 	}
 
 #ifdef CONFIG_LOCKDEP
 	acquire_freeze_lock(sb, level, !wait, _RET_IP_);
 #endif
 	percpu_counter_inc(&sb->s_writers.counter[level-1]);
-	/*
-	 * Make sure counter is updated before we check for frozen.
-	 * freeze_super() first sets frozen and then checks the counter.
-	 */
-	smp_mb();
-	if (unlikely(sb->s_writers.frozen >= level)) {
-		__sb_end_write(sb, level);
-		goto retry;
-	}
+	rcu_read_unlock();
 	return 1;
 }
 EXPORT_SYMBOL(__sb_start_write);
@@ -1312,7 +1310,13 @@ int freeze_super(struct super_block *sb)
 
 	/* From now on, no new normal writers can start */
 	sb->s_writers.frozen = SB_FREEZE_WRITE;
-	smp_wmb();
+	/*
+	 * After we synchronize_rcu(), we have ensured that everyone
+	 * who reads sb->s_writers.frozen under rcu_read_lock() can
+	 * now see our update.  This pretty much means that
+	 * __sb_start_write() will not allow any new writers.
+	 */
+	synchronize_rcu();
 
 	/* Release s_umount to preserve sb_start_write -> s_umount ordering */
 	up_write(&sb->s_umount);
@@ -1322,7 +1326,7 @@ int freeze_super(struct super_block *sb)
 	/* Now we go and block page faults... */
 	down_write(&sb->s_umount);
 	sb->s_writers.frozen = SB_FREEZE_PAGEFAULT;
-	smp_wmb();
+	synchronize_rcu();
 
 	sb_wait_write(sb, SB_FREEZE_PAGEFAULT);
 
@@ -1331,7 +1335,7 @@ int freeze_super(struct super_block *sb)
 
 	/* Now wait for internal filesystem counter */
 	sb->s_writers.frozen = SB_FREEZE_FS;
-	smp_wmb();
+	synchronize_rcu();
 	sb_wait_write(sb, SB_FREEZE_FS);
 
 	if (sb->s_op->freeze_fs) {
@@ -1340,7 +1344,7 @@ int freeze_super(struct super_block *sb)
 			printk(KERN_ERR
 				"VFS:Filesystem freeze failed\n");
 			sb->s_writers.frozen = SB_UNFROZEN;
-			smp_wmb();
+			synchronize_rcu();
 			wake_up(&sb->s_writers.wait_unfrozen);
 			deactivate_locked_super(sb);
 			return ret;
@@ -1387,7 +1391,7 @@ int thaw_super(struct super_block *sb)
 
 out:
 	sb->s_writers.frozen = SB_UNFROZEN;
-	smp_wmb();
+	synchronize_rcu();
 	wake_up(&sb->s_writers.wait_unfrozen);
 	deactivate_locked_super(sb);
 
_
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/

[RFC][PATCH 1/2] fs: use RCU for free_super() vs. __sb_start_write()

[Dave Hansen]

Currently, __sb_start_write() and freeze_super() can race with
each other.  __sb_start_write() uses a smp_mb() to ensure that
freeze_super() can see its write to sb->s_writers.counter and
that it can see freeze_super()'s update to sb->s_writers.frozen.
This all seems to work fine.

But, this smp_mb() makes __sb_start_write() the single hottest
function in the kernel if I sit in a loop and do tiny write()s to
tmpfs over and over.  This is on a very small 2-core system, so
it will only get worse on larger systems.

This _seems_ like an ideal case for RCU.  __sb_start_write() is
the RCU read-side and is in a very fast, performance-sensitive
path.  freeze_super() is the RCU writer and is in an extremely
rare non-performance-sensitive path.

Instead of doing and smp_wmb() in __sb_start_write(), we do
rcu_read_lock().  This ensures that a CPU doing freeze_super()
can not proceed past its synchronize_rcu() until the grace
period has ended and the 's_writers.frozen = SB_FREEZE_WRITE'
is visible to __sb_start_write().

One question here: Does the work that __sb_start_write() does in
a previous grace period becomes visible to freeze_super() after
its call to synchronize_rcu()?  It _seems_ like it should, but it
seems backwards to me since __sb_start_write() is the "reader" in
this case.

This patch increases the number of writes/second that I can do
by 10.4%.

Does anybody see any holes with this?

Cc: Jan Kara <jack@suse.cz>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>

---

 b/fs/super.c |   38 +++++++++++++++++++++-----------------
 1 file changed, 21 insertions(+), 17 deletions(-)

diff -puN fs/super.c~rcu-__sb_start_write fs/super.c
--- a/fs/super.c~rcu-__sb_start_write	2015-06-19 14:50:53.081869092 -0700
+++ b/fs/super.c	2015-06-19 15:19:03.000473047 -0700
@@ -1190,27 +1190,25 @@ static void acquire_freeze_lock(struct s
  */
 int __sb_start_write(struct super_block *sb, int level, bool wait)
 {
-retry:
-	if (unlikely(sb->s_writers.frozen >= level)) {
+	/*
+	 * RCU keeps freeze_super() from proceeding
+	 * while we are in here.
+	 */
+	rcu_read_lock();
+	while (unlikely(sb->s_writers.frozen >= level)) {
+		rcu_read_unlock();
 		if (!wait)
-			return 0;
+	       		return 0;
 		wait_event(sb->s_writers.wait_unfrozen,
 			   sb->s_writers.frozen < level);
+		rcu_read_lock();
 	}
 
 #ifdef CONFIG_LOCKDEP
 	acquire_freeze_lock(sb, level, !wait, _RET_IP_);
 #endif
 	percpu_counter_inc(&sb->s_writers.counter[level-1]);
-	/*
-	 * Make sure counter is updated before we check for frozen.
-	 * freeze_super() first sets frozen and then checks the counter.
-	 */
-	smp_mb();
-	if (unlikely(sb->s_writers.frozen >= level)) {
-		__sb_end_write(sb, level);
-		goto retry;
-	}
+	rcu_read_unlock();
 	return 1;
 }
 EXPORT_SYMBOL(__sb_start_write);
@@ -1312,7 +1310,13 @@ int freeze_super(struct super_block *sb)
 
 	/* From now on, no new normal writers can start */
 	sb->s_writers.frozen = SB_FREEZE_WRITE;
-	smp_wmb();
+	/*
+	 * After we synchronize_rcu(), we have ensured that everyone
+	 * who reads sb->s_writers.frozen under rcu_read_lock() can
+	 * now see our update.  This pretty much means that
+	 * __sb_start_write() will not allow any new writers.
+	 */
+	synchronize_rcu();
 
 	/* Release s_umount to preserve sb_start_write -> s_umount ordering */
 	up_write(&sb->s_umount);
@@ -1322,7 +1326,7 @@ int freeze_super(struct super_block *sb)
 	/* Now we go and block page faults... */
 	down_write(&sb->s_umount);
 	sb->s_writers.frozen = SB_FREEZE_PAGEFAULT;
-	smp_wmb();
+	synchronize_rcu();
 
 	sb_wait_write(sb, SB_FREEZE_PAGEFAULT);
 
@@ -1331,7 +1335,7 @@ int freeze_super(struct super_block *sb)
 
 	/* Now wait for internal filesystem counter */
 	sb->s_writers.frozen = SB_FREEZE_FS;
-	smp_wmb();
+	synchronize_rcu();
 	sb_wait_write(sb, SB_FREEZE_FS);
 
 	if (sb->s_op->freeze_fs) {
@@ -1340,7 +1344,7 @@ int freeze_super(struct super_block *sb)
 			printk(KERN_ERR
 				"VFS:Filesystem freeze failed\n");
 			sb->s_writers.frozen = SB_UNFROZEN;
-			smp_wmb();
+			synchronize_rcu();
 			wake_up(&sb->s_writers.wait_unfrozen);
 			deactivate_locked_super(sb);
 			return ret;
@@ -1387,7 +1391,7 @@ int thaw_super(struct super_block *sb)
 
 out:
 	sb->s_writers.frozen = SB_UNFROZEN;
-	smp_wmb();
+	synchronize_rcu();
 	wake_up(&sb->s_writers.wait_unfrozen);
 	deactivate_locked_super(sb);
 
_
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in

[RFC][PATCH 2/2] fs: conditionally do memory barrier in __sb_end_write()

[Dave Hansen]

If I sit in a loop and do write()s to small tmpfs files,
__sb_end_write() is third-hottest kernel function due to its
smp_mb().

The stated purpose for the smp_mb() in __sb_end_write() is to
ensure "s_writers are updated before we wake up waiters".  We
only wake up waiters if waitqueue_active(), but we do the
smp_mb() unconditionally.

It seems like we should be able to avoid it unless we are
actually doing the wake_up().

Cc: Jan Kara <jack@suse.cz>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>

---

 b/fs/super.c |   13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff -puN fs/super.c~selectively-do-barriers-in-__sb_end_write fs/super.c
--- a/fs/super.c~selectively-do-barriers-in-__sb_end_write	2015-06-19 15:20:37.953726659 -0700
+++ b/fs/super.c	2015-06-19 15:20:37.956726794 -0700
@@ -1147,13 +1147,14 @@ out:
 void __sb_end_write(struct super_block *sb, int level)
 {
 	percpu_counter_dec(&sb->s_writers.counter[level-1]);
-	/*
-	 * Make sure s_writers are updated before we wake up waiters in
-	 * freeze_super().
-	 */
-	smp_mb();
-	if (waitqueue_active(&sb->s_writers.wait))
+	if (waitqueue_active(&sb->s_writers.wait)) {
+		/*
+		 * Make sure other CPUs can see our s_writers update
+		 * before we wake up waiters in freeze_super().
+		 */
+		smp_mb();
 		wake_up(&sb->s_writers.wait);
+	}
 	rwsem_release(&sb->s_writers.lock_map[level-1], 1, _RET_IP_);
 }
 EXPORT_SYMBOL(__sb_end_write);
_
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
Please read the FAQ at  http://www.tux.org/lkml/

[RFC][PATCH 2/2] fs: conditionally do memory barrier in __sb_end_write()

[Dave Hansen]

If I sit in a loop and do write()s to small tmpfs files,
__sb_end_write() is third-hottest kernel function due to its
smp_mb().

The stated purpose for the smp_mb() in __sb_end_write() is to
ensure "s_writers are updated before we wake up waiters".  We
only wake up waiters if waitqueue_active(), but we do the
smp_mb() unconditionally.

It seems like we should be able to avoid it unless we are
actually doing the wake_up().

Cc: Jan Kara <jack@suse.cz>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>

---

 b/fs/super.c |   13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff -puN fs/super.c~selectively-do-barriers-in-__sb_end_write fs/super.c
--- a/fs/super.c~selectively-do-barriers-in-__sb_end_write	2015-06-19 15:20:37.953726659 -0700
+++ b/fs/super.c	2015-06-19 15:20:37.956726794 -0700
@@ -1147,13 +1147,14 @@ out:
 void __sb_end_write(struct super_block *sb, int level)
 {
 	percpu_counter_dec(&sb->s_writers.counter[level-1]);
-	/*
-	 * Make sure s_writers are updated before we wake up waiters in
-	 * freeze_super().
-	 */
-	smp_mb();
-	if (waitqueue_active(&sb->s_writers.wait))
+	if (waitqueue_active(&sb->s_writers.wait)) {
+		/*
+		 * Make sure other CPUs can see our s_writers update
+		 * before we wake up waiters in freeze_super().
+		 */
+		smp_mb();
 		wake_up(&sb->s_writers.wait);
+	}
 	rwsem_release(&sb->s_writers.lock_map[level-1], 1, _RET_IP_);
 }
 EXPORT_SYMBOL(__sb_end_write);
_
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in

Re: [RFC][PATCH 2/2] fs: conditionally do memory barrier in __sb_end_write()

[Jan Kara]

On Fri 19-06-15 15:32:23, Dave Hansen wrote:
> If I sit in a loop and do write()s to small tmpfs files,
> __sb_end_write() is third-hottest kernel function due to its
> smp_mb().
> 
> The stated purpose for the smp_mb() in __sb_end_write() is to
> ensure "s_writers are updated before we wake up waiters".  We
> only wake up waiters if waitqueue_active(), but we do the
> smp_mb() unconditionally.
> 
> It seems like we should be able to avoid it unless we are
> actually doing the wake_up().
...
> diff -puN fs/super.c~selectively-do-barriers-in-__sb_end_write fs/super.c
> --- a/fs/super.c~selectively-do-barriers-in-__sb_end_write	2015-06-19 15:20:37.953726659 -0700
> +++ b/fs/super.c	2015-06-19 15:20:37.956726794 -0700
> @@ -1147,13 +1147,14 @@ out:
>  void __sb_end_write(struct super_block *sb, int level)
>  {
>  	percpu_counter_dec(&sb->s_writers.counter[level-1]);
> -	/*
> -	 * Make sure s_writers are updated before we wake up waiters in
> -	 * freeze_super().
> -	 */
> -	smp_mb();
> -	if (waitqueue_active(&sb->s_writers.wait))
> +	if (waitqueue_active(&sb->s_writers.wait)) {
> +		/*
> +		 * Make sure other CPUs can see our s_writers update
> +		 * before we wake up waiters in freeze_super().
> +		 */
> +		smp_mb();

I think this is actually wrong. The barrier has to be before the
waitqueue_active() check. Otherwise that read can be reordered before the
percpu counter increment and a race window opens...

But we could make things faster by something like:

__sb_end_write()
	rcu_read_lock();
	percpu_counter_dec(&sb->s_writers.counter[level-1]);
	if (unlikely(sb->s_writers.frozen >= level))
		wake_up(&sb->s_writers.wait);
	rcu_read_unlock();

So the synchronize_rcu() calls you've added in the first patch will make
sure that all __sb_end_write() calls after we've started the freeze
procedure will end up calling wake_up() and so the process waiting in
sb_wait_write() will be woken as necessary. But please add a detailed
comment about the synchronization because its tricky and uncommon...

								Honza


>  		wake_up(&sb->s_writers.wait);
> +	}
>  	rwsem_release(&sb->s_writers.lock_map[level-1], 1, _RET_IP_);
>  }
>  EXPORT_SYMBOL(__sb_end_write);
> _
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR

Re: [RFC][PATCH 1/2] fs: use RCU for free_super() vs. __sb_start_write()

[Jan Kara]

On Fri 19-06-15 15:32:23, Dave Hansen wrote:
> 
> Currently, __sb_start_write() and freeze_super() can race with
> each other.  __sb_start_write() uses a smp_mb() to ensure that
> freeze_super() can see its write to sb->s_writers.counter and
> that it can see freeze_super()'s update to sb->s_writers.frozen.
> This all seems to work fine.
> 
> But, this smp_mb() makes __sb_start_write() the single hottest
> function in the kernel if I sit in a loop and do tiny write()s to
> tmpfs over and over.  This is on a very small 2-core system, so
> it will only get worse on larger systems.
> 
> This _seems_ like an ideal case for RCU.  __sb_start_write() is
> the RCU read-side and is in a very fast, performance-sensitive
> path.  freeze_super() is the RCU writer and is in an extremely
> rare non-performance-sensitive path.
> 
> Instead of doing and smp_wmb() in __sb_start_write(), we do
> rcu_read_lock().  This ensures that a CPU doing freeze_super()
> can not proceed past its synchronize_rcu() until the grace
> period has ended and the 's_writers.frozen = SB_FREEZE_WRITE'
> is visible to __sb_start_write().
> 
> One question here: Does the work that __sb_start_write() does in
> a previous grace period becomes visible to freeze_super() after
> its call to synchronize_rcu()?  It _seems_ like it should, but it
> seems backwards to me since __sb_start_write() is the "reader" in
> this case.
> 
> This patch increases the number of writes/second that I can do
> by 10.4%.
> 
> Does anybody see any holes with this?

Nice speed up and looks good to me. Just one question below.

> @@ -1340,7 +1344,7 @@ int freeze_super(struct super_block *sb)
>  			printk(KERN_ERR
>  				"VFS:Filesystem freeze failed\n");
>  			sb->s_writers.frozen = SB_UNFROZEN;
> -			smp_wmb();
> +			synchronize_rcu();

Do we really need synchronize_rcu() here? We just need to make sure write
to sb->s_writers.frozen happens before we start waking processes...

>  			wake_up(&sb->s_writers.wait_unfrozen);
>  			deactivate_locked_super(sb);
>  			return ret;
> @@ -1387,7 +1391,7 @@ int thaw_super(struct super_block *sb)
>  
>  out:
>  	sb->s_writers.frozen = SB_UNFROZEN;
> -	smp_wmb();
> +	synchronize_rcu();
>  	wake_up(&sb->s_writers.wait_unfrozen);

And here as well...

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR

Re: [RFC][PATCH 1/2] fs: use RCU for free_super() vs. __sb_start_write()

[Dave Hansen]

On 06/23/2015 04:09 AM, Jan Kara wrote:
>> @@ -1340,7 +1344,7 @@ int freeze_super(struct super_block *sb)
>>  			printk(KERN_ERR
>>  				"VFS:Filesystem freeze failed\n");
>>  			sb->s_writers.frozen = SB_UNFROZEN;
>> -			smp_wmb();
>> +			synchronize_rcu();
> 
> Do we really need synchronize_rcu() here? We just need to make sure write
> to sb->s_writers.frozen happens before we start waking processes...

I don't think it is necessary.  We only need to be concerned in practice
if someone could be inside a critical section when we are executing
this.  I *think* the only case that we have that really matters will be
taken care of by the _first_ synchronize_rcu().

It's definitely worth adding a comment.