[Ocfs2-devel] [PATCH 3/3] ocfs2: drop and retake orphan_dir.i_mutex in ocfs2_recover_orphans

[Ocfs2-devel] [PATCH 3/3] ocfs2: drop and retake orphan_dir.i_mutex in ocfs2_recover_orphans

[Wengang Wang]

drop and retake orphan_dir.i_mutex in ocfs2_recover_orphans.

Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
---
 fs/ocfs2/journal.c |   40 ++++++++++++++++++++++++++++++++++------
 1 files changed, 34 insertions(+), 6 deletions(-)

diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
index 7994823..e3b8340 100644
--- a/fs/ocfs2/journal.c
+++ b/fs/ocfs2/journal.c
@@ -56,6 +56,7 @@
 DEFINE_SPINLOCK(trans_inc_lock);
 
 #define ORPHAN_SCAN_SCHEDULE_TIMEOUT 300000
+#define ORPHAN_SCAN_ITEMS_PER_LOOP 200
 
 static int ocfs2_force_read_journal(struct inode *inode);
 static int ocfs2_recover_node(struct ocfs2_super *osb,
@@ -1949,6 +1950,8 @@ void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
 struct ocfs2_orphan_filldir_priv {
 	struct inode		*head;
 	struct ocfs2_super	*osb;
+	int			count;
+	int			for_live:1;
 };
 
 static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
@@ -1957,6 +1960,9 @@ static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
 	struct ocfs2_orphan_filldir_priv *p = priv;
 	struct inode *iter;
 
+	if (p->for_live)
+		BUG_ON(p->count > ORPHAN_SCAN_ITEMS_PER_LOOP);
+
 	if (name_len == 1 && !strncmp(".", name, 1))
 		return 0;
 	if (name_len == 2 && !strncmp("..", name, 2))
@@ -1974,20 +1980,38 @@ static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
 	OCFS2_I(iter)->ip_next_orphan = p->head;
 	p->head = iter;
 
+	/*
+	 * we fill at most ORPHAN_SCAN_ITEMS_PER_LOOP direntry in a loop with
+	 * the orphan_dir.i_mutex locked.
+	 */
+	if (p->for_live) {
+		if (++p->count == ORPHAN_SCAN_ITEMS_PER_LOOP)
+			return -EAGAIN;
+	}
 	return 0;
 }
 
 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
 			       int slot,
-			       struct inode **head)
+			       struct inode **head,
+			       loff_t *pos)
 {
 	int status;
 	struct inode *orphan_dir_inode = NULL;
 	struct ocfs2_orphan_filldir_priv priv;
-	loff_t pos = 0;
+	unsigned int node_num;
 
 	priv.osb = osb;
 	priv.head = *head;
+	priv.count = 0;
+
+	spin_lock(&osb->osb_lock);
+	status = ocfs2_slot_to_node_num_locked(osb, slot, &node_num);
+	spin_unlock(&osb->osb_lock);
+	if (!status && node_num == osb->node_num)
+		priv.for_live = true;
+	else
+		priv.for_live = false;
 
 	orphan_dir_inode = ocfs2_get_system_file_inode(osb,
 						       ORPHAN_DIR_SYSTEM_INODE,
@@ -2005,9 +2029,9 @@ static int ocfs2_queue_orphans(struct ocfs2_super *osb,
 		goto out;
 	}
 
-	status = ocfs2_dir_foreach(orphan_dir_inode, &pos, &priv,
+	status = ocfs2_dir_foreach(orphan_dir_inode, pos, &priv,
 				   ocfs2_orphan_filldir);
-	if (status) {
+	if (status && status != -EAGAIN) {
 		mlog_errno(status);
 		goto out_cluster;
 	}
@@ -2083,16 +2107,18 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
 	struct inode *inode = NULL;
 	struct inode *iter;
 	struct ocfs2_inode_info *oi;
+	loff_t pos = 0;
 
 	trace_ocfs2_recover_orphans(slot);
 
+cnt_scan:
 	ocfs2_mark_recovering_orphan_dir(osb, slot);
-	ret = ocfs2_queue_orphans(osb, slot, &inode);
+	ret = ocfs2_queue_orphans(osb, slot, &inode, &pos);
 	ocfs2_clear_recovering_orphan_dir(osb, slot);
 
 	/* Error here should be noted, but we want to continue with as
 	 * many queued inodes as we've got. */
-	if (ret)
+	if (ret && ret != -EAGAIN)
 		mlog_errno(ret);
 
 	while (inode) {
@@ -2119,6 +2145,8 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
 		inode = iter;
 	}
 
+	if (ret == -EAGAIN)
+		goto cnt_scan;
 	return ret;
 }
 
-- 
1.7.5.2

[Ocfs2-devel] [PATCH 3/3] ocfs2: drop and retake orphan_dir.i_mutex in ocfs2_recover_orphans

[Sunil Mushran]

This is too hacky for mainline. May work as a temp fix only.

This assumes that we only need to recover orphans in our own slot.
If that is the case, then if nothing else, we can remove the orphan
scan lock. We added the lock so as to ensure all nodes get a chance
to replay it. If all nodes are doing there own, they why do we need
the lock.

On 07/05/2011 10:11 PM, Wengang Wang wrote:
> drop and retake orphan_dir.i_mutex in ocfs2_recover_orphans.
>
> Signed-off-by: Wengang Wang<wen.gang.wang@oracle.com>
> ---
>   fs/ocfs2/journal.c |   40 ++++++++++++++++++++++++++++++++++------
>   1 files changed, 34 insertions(+), 6 deletions(-)
>
> diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
> index 7994823..e3b8340 100644
> --- a/fs/ocfs2/journal.c
> +++ b/fs/ocfs2/journal.c
> @@ -56,6 +56,7 @@
>   DEFINE_SPINLOCK(trans_inc_lock);
>
>   #define ORPHAN_SCAN_SCHEDULE_TIMEOUT 300000
> +#define ORPHAN_SCAN_ITEMS_PER_LOOP 200
>
>   static int ocfs2_force_read_journal(struct inode *inode);
>   static int ocfs2_recover_node(struct ocfs2_super *osb,
> @@ -1949,6 +1950,8 @@ void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
>   struct ocfs2_orphan_filldir_priv {
>   	struct inode		*head;
>   	struct ocfs2_super	*osb;
> +	int			count;
> +	int			for_live:1;
>   };
>
>   static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
> @@ -1957,6 +1960,9 @@ static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
>   	struct ocfs2_orphan_filldir_priv *p = priv;
>   	struct inode *iter;
>
> +	if (p->for_live)
> +		BUG_ON(p->count>  ORPHAN_SCAN_ITEMS_PER_LOOP);
> +
>   	if (name_len == 1&&  !strncmp(".", name, 1))
>   		return 0;
>   	if (name_len == 2&&  !strncmp("..", name, 2))
> @@ -1974,20 +1980,38 @@ static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
>   	OCFS2_I(iter)->ip_next_orphan = p->head;
>   	p->head = iter;
>
> +	/*
> +	 * we fill at most ORPHAN_SCAN_ITEMS_PER_LOOP direntry in a loop with
> +	 * the orphan_dir.i_mutex locked.
> +	 */
> +	if (p->for_live) {
> +		if (++p->count == ORPHAN_SCAN_ITEMS_PER_LOOP)
> +			return -EAGAIN;
> +	}
>   	return 0;
>   }
>
>   static int ocfs2_queue_orphans(struct ocfs2_super *osb,
>   			       int slot,
> -			       struct inode **head)
> +			       struct inode **head,
> +			       loff_t *pos)
>   {
>   	int status;
>   	struct inode *orphan_dir_inode = NULL;
>   	struct ocfs2_orphan_filldir_priv priv;
> -	loff_t pos = 0;
> +	unsigned int node_num;
>
>   	priv.osb = osb;
>   	priv.head = *head;
> +	priv.count = 0;
> +
> +	spin_lock(&osb->osb_lock);
> +	status = ocfs2_slot_to_node_num_locked(osb, slot,&node_num);
> +	spin_unlock(&osb->osb_lock);
> +	if (!status&&  node_num == osb->node_num)
> +		priv.for_live = true;
> +	else
> +		priv.for_live = false;

You cannot return (save) pos and yet lookup the slotnum again. That's
not the right order. Too hackish. All these functions will need to be
re-prototyped if your approach is what we use.

>
>   	orphan_dir_inode = ocfs2_get_system_file_inode(osb,
>   						       ORPHAN_DIR_SYSTEM_INODE,
> @@ -2005,9 +2029,9 @@ static int ocfs2_queue_orphans(struct ocfs2_super *osb,
>   		goto out;
>   	}
>
> -	status = ocfs2_dir_foreach(orphan_dir_inode,&pos,&priv,
> +	status = ocfs2_dir_foreach(orphan_dir_inode, pos,&priv,
>   				   ocfs2_orphan_filldir);
> -	if (status) {
> +	if (status&&  status != -EAGAIN) {
>   		mlog_errno(status);
>   		goto out_cluster;
>   	}
> @@ -2083,16 +2107,18 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
>   	struct inode *inode = NULL;
>   	struct inode *iter;
>   	struct ocfs2_inode_info *oi;
> +	loff_t pos = 0;
>
>   	trace_ocfs2_recover_orphans(slot);
>
> +cnt_scan:
>   	ocfs2_mark_recovering_orphan_dir(osb, slot);
> -	ret = ocfs2_queue_orphans(osb, slot,&inode);
> +	ret = ocfs2_queue_orphans(osb, slot,&inode,&pos);
>   	ocfs2_clear_recovering_orphan_dir(osb, slot);

The slot is being marked and cleared multiple times. I am unsure what the
ramifications are, but it does not look right.

>   	/* Error here should be noted, but we want to continue with as
>   	 * many queued inodes as we've got. */
> -	if (ret)
> +	if (ret&&  ret != -EAGAIN)
>   		mlog_errno(ret);
>
>   	while (inode) {
> @@ -2119,6 +2145,8 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
>   		inode = iter;
>   	}
>
> +	if (ret == -EAGAIN)
> +		goto cnt_scan;
>   	return ret;
>   }
>

[Ocfs2-devel] [PATCH 3/3] ocfs2: drop and retake orphan_dir.i_mutex in ocfs2_recover_orphans

[Wengang Wang]

On 11-07-07 14:48, Sunil Mushran wrote:
> This is too hacky for mainline. May work as a temp fix only.

:D

> 
> This assumes that we only need to recover orphans in our own slot.
> If that is the case, then if nothing else, we can remove the orphan
> scan lock. We added the lock so as to ensure all nodes get a chance
> to replay it. If all nodes are doing there own, they why do we need
> the lock.

Actually, for live nodes, yes we only recover our own slot. For the
dead slots, we need the orphan scan lock.

> 
> On 07/05/2011 10:11 PM, Wengang Wang wrote:
> >drop and retake orphan_dir.i_mutex in ocfs2_recover_orphans.
> >
> >Signed-off-by: Wengang Wang<wen.gang.wang@oracle.com>
> >---
> >  fs/ocfs2/journal.c |   40 ++++++++++++++++++++++++++++++++++------
> >  1 files changed, 34 insertions(+), 6 deletions(-)
> >
> >diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
> >index 7994823..e3b8340 100644
> >--- a/fs/ocfs2/journal.c
> >+++ b/fs/ocfs2/journal.c
> >@@ -56,6 +56,7 @@
> >  DEFINE_SPINLOCK(trans_inc_lock);
> >
> >  #define ORPHAN_SCAN_SCHEDULE_TIMEOUT 300000
> >+#define ORPHAN_SCAN_ITEMS_PER_LOOP 200
> >
> >  static int ocfs2_force_read_journal(struct inode *inode);
> >  static int ocfs2_recover_node(struct ocfs2_super *osb,
> >@@ -1949,6 +1950,8 @@ void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
> >  struct ocfs2_orphan_filldir_priv {
> >  	struct inode		*head;
> >  	struct ocfs2_super	*osb;
> >+	int			count;
> >+	int			for_live:1;
> >  };
> >
> >  static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
> >@@ -1957,6 +1960,9 @@ static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
> >  	struct ocfs2_orphan_filldir_priv *p = priv;
> >  	struct inode *iter;
> >
> >+	if (p->for_live)
> >+		BUG_ON(p->count>  ORPHAN_SCAN_ITEMS_PER_LOOP);
> >+
> >  	if (name_len == 1&&  !strncmp(".", name, 1))
> >  		return 0;
> >  	if (name_len == 2&&  !strncmp("..", name, 2))
> >@@ -1974,20 +1980,38 @@ static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
> >  	OCFS2_I(iter)->ip_next_orphan = p->head;
> >  	p->head = iter;
> >
> >+	/*
> >+	 * we fill at most ORPHAN_SCAN_ITEMS_PER_LOOP direntry in a loop with
> >+	 * the orphan_dir.i_mutex locked.
> >+	 */
> >+	if (p->for_live) {
> >+		if (++p->count == ORPHAN_SCAN_ITEMS_PER_LOOP)
> >+			return -EAGAIN;
> >+	}
> >  	return 0;
> >  }
> >
> >  static int ocfs2_queue_orphans(struct ocfs2_super *osb,
> >  			       int slot,
> >-			       struct inode **head)
> >+			       struct inode **head,
> >+			       loff_t *pos)
> >  {
> >  	int status;
> >  	struct inode *orphan_dir_inode = NULL;
> >  	struct ocfs2_orphan_filldir_priv priv;
> >-	loff_t pos = 0;
> >+	unsigned int node_num;
> >
> >  	priv.osb = osb;
> >  	priv.head = *head;
> >+	priv.count = 0;
> >+
> >+	spin_lock(&osb->osb_lock);
> >+	status = ocfs2_slot_to_node_num_locked(osb, slot,&node_num);
> >+	spin_unlock(&osb->osb_lock);
> >+	if (!status&&  node_num == osb->node_num)
> >+		priv.for_live = true;
> >+	else
> >+		priv.for_live = false;
> 
> You cannot return (save) pos and yet lookup the slotnum again. That's
> not the right order. Too hackish. All these functions will need to be
> re-prototyped if your approach is what we use.

Yes, I also think it's hackish.
But I don't have a better fix so far. While though it's hackish, it works.
It is just like ocfs2_readdir(). If a new direntry is added before the
position, we just leave there for next run; If one is added after the
possition, we take it; If a diretry is removed, I means the file/inode
is freed, OK; It's safe.
If we restart from offset 0, in case all the inodes in the first 200 or
whatever other limit are not freed and won't be freed(say still opened),
we are in bad loops. 

> 
> >
> >  	orphan_dir_inode = ocfs2_get_system_file_inode(osb,
> >  						       ORPHAN_DIR_SYSTEM_INODE,
> >@@ -2005,9 +2029,9 @@ static int ocfs2_queue_orphans(struct ocfs2_super *osb,
> >  		goto out;
> >  	}
> >
> >-	status = ocfs2_dir_foreach(orphan_dir_inode,&pos,&priv,
> >+	status = ocfs2_dir_foreach(orphan_dir_inode, pos,&priv,
> >  				   ocfs2_orphan_filldir);
> >-	if (status) {
> >+	if (status&&  status != -EAGAIN) {
> >  		mlog_errno(status);
> >  		goto out_cluster;
> >  	}
> >@@ -2083,16 +2107,18 @@ static int ocfs2_recover_orphans(struct ocfs2_super *osb,
> >  	struct inode *inode = NULL;
> >  	struct inode *iter;
> >  	struct ocfs2_inode_info *oi;
> >+	loff_t pos = 0;
> >
> >  	trace_ocfs2_recover_orphans(slot);
> >
> >+cnt_scan:
> >  	ocfs2_mark_recovering_orphan_dir(osb, slot);
> >-	ret = ocfs2_queue_orphans(osb, slot,&inode);
> >+	ret = ocfs2_queue_orphans(osb, slot,&inode,&pos);
> >  	ocfs2_clear_recovering_orphan_dir(osb, slot);
> 
> The slot is being marked and cleared multiple times. I am unsure what the
> ramifications are, but it does not look right.

I will look at this more...

thanks,
wengang.

[Ocfs2-devel] [PATCH 3/3] ocfs2: drop and retake orphan_dir.i_mutex in ocfs2_recover_orphans

[Wengang Wang]

On 11-07-08 10:28, Wengang Wang wrote:
> On 11-07-07 14:48, Sunil Mushran wrote:
> > This is too hacky for mainline. May work as a temp fix only.

<snip>

> > >  	trace_ocfs2_recover_orphans(slot);
> > >
> > >+cnt_scan:
> > >  	ocfs2_mark_recovering_orphan_dir(osb, slot);
> > >-	ret = ocfs2_queue_orphans(osb, slot,&inode);
> > >+	ret = ocfs2_queue_orphans(osb, slot,&inode,&pos);
> > >  	ocfs2_clear_recovering_orphan_dir(osb, slot);
> > 
> > The slot is being marked and cleared multiple times. I am unsure what the
> > ramifications are, but it does not look right.
> 
> I will look at this more...

Per my understand, ocfs2_mark_recovering_orphan_dir() is used to avoid
dead lock only.

ABBA lock.
A: inode lock(cluster lock) on inode to delete
B: inode lock on orphan_dir inode

the ocfs2_delete_inode() path takes in AB.
the ocfs2_recover_orphans() path takes in BA on a different node.

Since AB are both required and then released within ocfs2_queue_orphans(), calling
ocfs2_mark_recovering_orphan_dir() multiple times is not a problem.

Opition?

thanks,
wengang.