[PATCH] gfs2: Stop using rhashtable_walk_peek

[PATCH] gfs2: Stop using rhashtable_walk_peek

[Andreas Gruenbacher]

Function rhashtable_walk_peek is problematic because there is no
guarantee that the glock previously returned still exists; when that key
is deleted, rhashtable_walk_peek can end up returning a different key,
which would cause an inconsistent glock dump.  So instead of using
rhashtable_walk_peek, keep track of the current glock in the seq file
iterator functions.

Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
---
 fs/gfs2/glock.c | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)

diff --git a/fs/gfs2/glock.c b/fs/gfs2/glock.c
index 82fb5583445c..f1fc353875d3 100644
--- a/fs/gfs2/glock.c
+++ b/fs/gfs2/glock.c
@@ -55,6 +55,7 @@ struct gfs2_glock_iter {
 	struct gfs2_sbd *sdp;		/* incore superblock           */
 	struct rhashtable_iter hti;	/* rhashtable iterator         */
 	struct gfs2_glock *gl;		/* current glock struct        */
+	bool gl_held;
 	loff_t last_pos;		/* last position               */
 };
 
@@ -1923,9 +1924,11 @@ void gfs2_glock_exit(void)
 
 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
 {
-	if (n == 0)
-		gi->gl = rhashtable_walk_peek(&gi->hti);
-	else {
+	if (n != 0 || !gi->gl) {
+		if (gi->gl_held) {
+			gfs2_glock_queue_put(gi->gl);
+			gi->gl_held = false;
+		}
 		gi->gl = rhashtable_walk_next(&gi->hti);
 		n--;
 	}
@@ -1988,7 +1991,10 @@ static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
 {
 	struct gfs2_glock_iter *gi = seq->private;
 
-	gi->gl = NULL;
+	if (gi->gl) {
+		lockref_get(&gi->gl->gl_lockref);
+		gi->gl_held = true;
+	}
 	rhashtable_walk_stop(&gi->hti);
 }
 
@@ -2061,6 +2067,7 @@ static int __gfs2_glocks_open(struct inode *inode, struct file *file,
 		 */
 		gi->last_pos = -1;
 		gi->gl = NULL;
+		gi->gl_held = false;
 		rhashtable_walk_enter(&gl_hash_table, &gi->hti);
 	}
 	return ret;
@@ -2076,7 +2083,8 @@ static int gfs2_glocks_release(struct inode *inode, struct file *file)
 	struct seq_file *seq = file->private_data;
 	struct gfs2_glock_iter *gi = seq->private;
 
-	gi->gl = NULL;
+	if (gi->gl_held)
+		gfs2_glock_put(gi->gl);
 	rhashtable_walk_exit(&gi->hti);
 	return seq_release_private(inode, file);
 }
-- 
2.14.3

[Cluster-devel] [PATCH] gfs2: Stop using rhashtable_walk_peek

[Andreas Gruenbacher]

Function rhashtable_walk_peek is problematic because there is no
guarantee that the glock previously returned still exists; when that key
is deleted, rhashtable_walk_peek can end up returning a different key,
which would cause an inconsistent glock dump.  So instead of using
rhashtable_walk_peek, keep track of the current glock in the seq file
iterator functions.

Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
---
 fs/gfs2/glock.c | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)

diff --git a/fs/gfs2/glock.c b/fs/gfs2/glock.c
index 82fb5583445c..f1fc353875d3 100644
--- a/fs/gfs2/glock.c
+++ b/fs/gfs2/glock.c
@@ -55,6 +55,7 @@ struct gfs2_glock_iter {
 	struct gfs2_sbd *sdp;		/* incore superblock           */
 	struct rhashtable_iter hti;	/* rhashtable iterator         */
 	struct gfs2_glock *gl;		/* current glock struct        */
+	bool gl_held;
 	loff_t last_pos;		/* last position               */
 };
 
@@ -1923,9 +1924,11 @@ void gfs2_glock_exit(void)
 
 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
 {
-	if (n == 0)
-		gi->gl = rhashtable_walk_peek(&gi->hti);
-	else {
+	if (n != 0 || !gi->gl) {
+		if (gi->gl_held) {
+			gfs2_glock_queue_put(gi->gl);
+			gi->gl_held = false;
+		}
 		gi->gl = rhashtable_walk_next(&gi->hti);
 		n--;
 	}
@@ -1988,7 +1991,10 @@ static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
 {
 	struct gfs2_glock_iter *gi = seq->private;
 
-	gi->gl = NULL;
+	if (gi->gl) {
+		lockref_get(&gi->gl->gl_lockref);
+		gi->gl_held = true;
+	}
 	rhashtable_walk_stop(&gi->hti);
 }
 
@@ -2061,6 +2067,7 @@ static int __gfs2_glocks_open(struct inode *inode, struct file *file,
 		 */
 		gi->last_pos = -1;
 		gi->gl = NULL;
+		gi->gl_held = false;
 		rhashtable_walk_enter(&gl_hash_table, &gi->hti);
 	}
 	return ret;
@@ -2076,7 +2083,8 @@ static int gfs2_glocks_release(struct inode *inode, struct file *file)
 	struct seq_file *seq = file->private_data;
 	struct gfs2_glock_iter *gi = seq->private;
 
-	gi->gl = NULL;
+	if (gi->gl_held)
+		gfs2_glock_put(gi->gl);
 	rhashtable_walk_exit(&gi->hti);
 	return seq_release_private(inode, file);
 }
-- 
2.14.3

Re: [PATCH] gfs2: Stop using rhashtable_walk_peek

[NeilBrown]

[-- Attachment #1: Type: text/plain, Size: 2913 bytes --]

On Wed, Mar 28 2018, Andreas Gruenbacher wrote:

> Function rhashtable_walk_peek is problematic because there is no
> guarantee that the glock previously returned still exists; when that key
> is deleted, rhashtable_walk_peek can end up returning a different key,
> which would cause an inconsistent glock dump.  So instead of using
> rhashtable_walk_peek, keep track of the current glock in the seq file
> iterator functions.
>
> Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
> ---
>  fs/gfs2/glock.c | 18 +++++++++++++-----
>  1 file changed, 13 insertions(+), 5 deletions(-)
>
> diff --git a/fs/gfs2/glock.c b/fs/gfs2/glock.c
> index 82fb5583445c..f1fc353875d3 100644
> --- a/fs/gfs2/glock.c
> +++ b/fs/gfs2/glock.c
> @@ -55,6 +55,7 @@ struct gfs2_glock_iter {
>  	struct gfs2_sbd *sdp;		/* incore superblock           */
>  	struct rhashtable_iter hti;	/* rhashtable iterator         */
>  	struct gfs2_glock *gl;		/* current glock struct        */
> +	bool gl_held;
>  	loff_t last_pos;		/* last position               */
>  };
>  
> @@ -1923,9 +1924,11 @@ void gfs2_glock_exit(void)
>  
>  static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
>  {
> -	if (n == 0)
> -		gi->gl = rhashtable_walk_peek(&gi->hti);
> -	else {
> +	if (n != 0 || !gi->gl) {
> +		if (gi->gl_held) {
> +			gfs2_glock_queue_put(gi->gl);
> +			gi->gl_held = false;
> +		}
>  		gi->gl = rhashtable_walk_next(&gi->hti);
>  		n--;
>  	}

Thank for this patch!
The above looks a bit fragile to me.
gfs2_glock_iter_next() (And hence gfs2_glock_seq_start()) will sometimes
exit with gl_held true, and sometimes with it false.
gfs2_glock_seq_stop() assumes that it is false.
Normally gfs2_glock_seq_next() will normally be called between these
two and will clear gl_held, but I don't think there is a hard guarantee
of that.
Maybe we should always 'put' gi->gl in iter_next if gl_held??

Thanks,
NeilBrown



> @@ -1988,7 +1991,10 @@ static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
>  {
>  	struct gfs2_glock_iter *gi = seq->private;
>  
> -	gi->gl = NULL;
> +	if (gi->gl) {
> +		lockref_get(&gi->gl->gl_lockref);
> +		gi->gl_held = true;
> +	}
>  	rhashtable_walk_stop(&gi->hti);
>  }
>  
> @@ -2061,6 +2067,7 @@ static int __gfs2_glocks_open(struct inode *inode, struct file *file,
>  		 */
>  		gi->last_pos = -1;
>  		gi->gl = NULL;
> +		gi->gl_held = false;
>  		rhashtable_walk_enter(&gl_hash_table, &gi->hti);
>  	}
>  	return ret;
> @@ -2076,7 +2083,8 @@ static int gfs2_glocks_release(struct inode *inode, struct file *file)
>  	struct seq_file *seq = file->private_data;
>  	struct gfs2_glock_iter *gi = seq->private;
>  
> -	gi->gl = NULL;
> +	if (gi->gl_held)
> +		gfs2_glock_put(gi->gl);
>  	rhashtable_walk_exit(&gi->hti);
>  	return seq_release_private(inode, file);
>  }
> -- 
> 2.14.3

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]

[Cluster-devel] [PATCH] gfs2: Stop using rhashtable_walk_peek

[NeilBrown]

On Wed, Mar 28 2018, Andreas Gruenbacher wrote:

> Function rhashtable_walk_peek is problematic because there is no
> guarantee that the glock previously returned still exists; when that key
> is deleted, rhashtable_walk_peek can end up returning a different key,
> which would cause an inconsistent glock dump.  So instead of using
> rhashtable_walk_peek, keep track of the current glock in the seq file
> iterator functions.
>
> Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
> ---
>  fs/gfs2/glock.c | 18 +++++++++++++-----
>  1 file changed, 13 insertions(+), 5 deletions(-)
>
> diff --git a/fs/gfs2/glock.c b/fs/gfs2/glock.c
> index 82fb5583445c..f1fc353875d3 100644
> --- a/fs/gfs2/glock.c
> +++ b/fs/gfs2/glock.c
> @@ -55,6 +55,7 @@ struct gfs2_glock_iter {
>  	struct gfs2_sbd *sdp;		/* incore superblock           */
>  	struct rhashtable_iter hti;	/* rhashtable iterator         */
>  	struct gfs2_glock *gl;		/* current glock struct        */
> +	bool gl_held;
>  	loff_t last_pos;		/* last position               */
>  };
>  
> @@ -1923,9 +1924,11 @@ void gfs2_glock_exit(void)
>  
>  static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
>  {
> -	if (n == 0)
> -		gi->gl = rhashtable_walk_peek(&gi->hti);
> -	else {
> +	if (n != 0 || !gi->gl) {
> +		if (gi->gl_held) {
> +			gfs2_glock_queue_put(gi->gl);
> +			gi->gl_held = false;
> +		}
>  		gi->gl = rhashtable_walk_next(&gi->hti);
>  		n--;
>  	}

Thank for this patch!
The above looks a bit fragile to me.
gfs2_glock_iter_next() (And hence gfs2_glock_seq_start()) will sometimes
exit with gl_held true, and sometimes with it false.
gfs2_glock_seq_stop() assumes that it is false.
Normally gfs2_glock_seq_next() will normally be called between these
two and will clear gl_held, but I don't think there is a hard guarantee
of that.
Maybe we should always 'put' gi->gl in iter_next if gl_held??

Thanks,
NeilBrown



> @@ -1988,7 +1991,10 @@ static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
>  {
>  	struct gfs2_glock_iter *gi = seq->private;
>  
> -	gi->gl = NULL;
> +	if (gi->gl) {
> +		lockref_get(&gi->gl->gl_lockref);
> +		gi->gl_held = true;
> +	}
>  	rhashtable_walk_stop(&gi->hti);
>  }
>  
> @@ -2061,6 +2067,7 @@ static int __gfs2_glocks_open(struct inode *inode, struct file *file,
>  		 */
>  		gi->last_pos = -1;
>  		gi->gl = NULL;
> +		gi->gl_held = false;
>  		rhashtable_walk_enter(&gl_hash_table, &gi->hti);
>  	}
>  	return ret;
> @@ -2076,7 +2083,8 @@ static int gfs2_glocks_release(struct inode *inode, struct file *file)
>  	struct seq_file *seq = file->private_data;
>  	struct gfs2_glock_iter *gi = seq->private;
>  
> -	gi->gl = NULL;
> +	if (gi->gl_held)
> +		gfs2_glock_put(gi->gl);
>  	rhashtable_walk_exit(&gi->hti);
>  	return seq_release_private(inode, file);
>  }
> -- 
> 2.14.3
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://listman.redhat.com/archives/cluster-devel/attachments/20180329/cd2b1142/attachment.sig>

Re: [PATCH] gfs2: Stop using rhashtable_walk_peek

[Andreas Gruenbacher]

On 28 March 2018 at 23:53, NeilBrown <neilb@suse.com> wrote:
> Thank for this patch!
> The above looks a bit fragile to me.
> gfs2_glock_iter_next() (And hence gfs2_glock_seq_start()) will sometimes
> exit with gl_held true, and sometimes with it false.
> gfs2_glock_seq_stop() assumes that it is false.
> Normally gfs2_glock_seq_next() will normally be called between these
> two and will clear gl_held, but I don't think there is a hard guarantee
> of that.
> Maybe we should always 'put' gi->gl in iter_next if gl_held??

Another problem is that gfs2_glock_seq_stop only does a lockref_get
which is not allowed because the lockref may have been marked dead in
the meantime. We'll have to take a reference on each lock in
gfs2_glock_iter_next instead. Revised patch to follow tomorrow.

Thanks,
Andreas

[Cluster-devel] [PATCH] gfs2: Stop using rhashtable_walk_peek

[Andreas Gruenbacher]

On 28 March 2018 at 23:53, NeilBrown <neilb@suse.com> wrote:
> Thank for this patch!
> The above looks a bit fragile to me.
> gfs2_glock_iter_next() (And hence gfs2_glock_seq_start()) will sometimes
> exit with gl_held true, and sometimes with it false.
> gfs2_glock_seq_stop() assumes that it is false.
> Normally gfs2_glock_seq_next() will normally be called between these
> two and will clear gl_held, but I don't think there is a hard guarantee
> of that.
> Maybe we should always 'put' gi->gl in iter_next if gl_held??

Another problem is that gfs2_glock_seq_stop only does a lockref_get
which is not allowed because the lockref may have been marked dead in
the meantime. We'll have to take a reference on each lock in
gfs2_glock_iter_next instead. Revised patch to follow tomorrow.

Thanks,
Andreas