ocfs2: reflink: fix slow unlink for refcounted file

When running ocfs2 test suite multiple nodes reflink stress test, for a 4 nodes cluster, every unlink() for refcounted file needs about 700s.

The slow unlink is caused by the contention of refcount tree lock since all nodes are unlink files using the same refcount tree. When the unlinking file have many extents(over 1600 in our test), most of the extents has refcounted flag set. In ocfs2_commit_truncate(), it will execute the following call trace for every extents. This means it needs get and released refcount tree lock about 1600 times. And when several nodes are do this at the same time, the performance will be very low.
.
ocfs2_remove_btree_range()
----ocfs2_lock_refcount_tree()
------ocfs2_refcount_lock()
--------__ocfs2_cluster_lock()

ocfs2_refcount_lock() is costly, move it to ocfs2_commit_truncate() to do lock/unlock once can improve a lot performance.

Signed-off-by Junxiao Bi <[email protected]>

---
  fs/ocfs2/alloc.c |   22 ++++++++++++----------
  1 file changed, 12 insertions(+), 10 deletions(-)

fs/ocfs2/alloc.c +12 -10

index 9d8fcf2..a764363 100644
@@ -5667,13 +5667,6 @@ int ocfs2_remove_btree_range(struct inode *inode,
  		BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
  			 OCFS2_HAS_REFCOUNT_FL));
 
-		ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
-					       &ref_tree, NULL);
-		if (ret) {
-			mlog_errno(ret);
-			goto bail;
-		}
-
  		ret = ocfs2_prepare_refcount_change_for_del(inode,
  							    refcount_loc,
  							    phys_blkno,

@@ -5755,9 +5748,6 @@ bail:
  	if (meta_ac)
  		ocfs2_free_alloc_context(meta_ac);
 
-	if (ref_tree)
-		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
-
  	return ret;
  }
 

@@ -7012,6 +7002,7 @@ int ocfs2_commit_truncate(struct ocfs2_super *osb,
  	u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
  	struct ocfs2_extent_tree et;
  	struct ocfs2_cached_dealloc_ctxt dealloc;
+	struct ocfs2_refcount_tree *ref_tree = NULL;
 
  	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
  	ocfs2_init_dealloc_ctxt(&dealloc);

@@ -7121,6 +7112,15 @@ start:
 
  	phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
 
+	if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
+		status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
+				&ref_tree, NULL);
+		if (status) {
+			mlog_errno(status);
+			goto bail;
+		}
+	}
+
  	status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
  					  phys_cpos, trunc_len, flags, &dealloc,
  					  refcount_loc);

@@ -7138,6 +7138,8 @@ start:
  	goto start;
 
  bail:
+	if (ref_tree)
+		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 
  	ocfs2_schedule_truncate_log_flush(osb, 1);
 

Will it be safe to do the moving?

There are three calls of ocfs2_remove_btree_range

1 7124 fs/ocfs2/alloc.c <<ocfs2_commit_truncate>> status = ocfs2_remove_btree_range(inode, &et, trunc_cpos, 2 4479 fs/ocfs2/dir.c <<ocfs2_dx_dir_truncate>>
ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0, 3 1805 fs/ocfs2/file.c <<ocfs2_remove_inode_range>> ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,

Obviously, the 2nd and the 3rd are not sub-routines of ocfs2_commit_truncate().

I think this moving is not safe.

thanks,
wengang

? 2014?07?25? 13:32, Junxiao Bi ??:

Quoted 77 lines

When running ocfs2 test suite multiple nodes reflink stress test, for a 4 nodes cluster, every unlink() for refcounted file needs about 700s.

The slow unlink is caused by the contention of refcount tree lock since all nodes are unlink files using the same refcount tree. When the unlinking file have many extents(over 1600 in our test), most of the extents has refcounted flag set. In ocfs2_commit_truncate(), it will execute the following call trace for every extents. This means it needs get and released refcount tree lock about 1600 times. And when several nodes are do this at the same time, the performance will be very low.
.
ocfs2_remove_btree_range()
----ocfs2_lock_refcount_tree()
------ocfs2_refcount_lock()
--------__ocfs2_cluster_lock()

ocfs2_refcount_lock() is costly, move it to ocfs2_commit_truncate() to do lock/unlock once can improve a lot performance.

Signed-off-by Junxiao Bi <[email protected]>

---
   fs/ocfs2/alloc.c |   22 ++++++++++++----------
   1 file changed, 12 insertions(+), 10 deletions(-)

fs/ocfs2/alloc.c +10 -10

index 9d8fcf2..a764363 100644
@@ -5667,13 +5667,6 @@ int ocfs2_remove_btree_range(struct inode *inode,
   		BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
   			 OCFS2_HAS_REFCOUNT_FL));
 
-		ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
-					       &ref_tree, NULL);
-		if (ret) {
-			mlog_errno(ret);
-			goto bail;
-		}
-
   		ret = ocfs2_prepare_refcount_change_for_del(inode,
   							    refcount_loc,
   							    phys_blkno,

@@ -5755,9 +5748,6 @@ bail:
   	if (meta_ac)
   		ocfs2_free_alloc_context(meta_ac);
 
-	if (ref_tree)
-		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
-
   	return ret;
   }
 

@@ -7012,6 +7002,7 @@ int ocfs2_commit_truncate(struct ocfs2_super *osb,
   	u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
   	struct ocfs2_extent_tree et;
   	struct ocfs2_cached_dealloc_ctxt dealloc;
+	struct ocfs2_refcount_tree *ref_tree = NULL;
 
   	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
   	ocfs2_init_dealloc_ctxt(&dealloc);

@@ -7121,6 +7112,15 @@ start:
 
   	phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
 
+	if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
+		status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
+				&ref_tree, NULL);
+		if (status) {
+			mlog_errno(status);
+			goto bail;
+		}
+	}
+
   	status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
   					  phys_cpos, trunc_len, flags, &dealloc,
   					  refcount_loc);

fs/ocfs2/alloc.c start: · line 7145

Quoted 8 lines

fs/ocfs2/alloc.c +2 -0

@@ -7138,6 +7138,8 @@ start:
   	goto start;
 
   bail:
+	if (ref_tree)
+		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 
   	ocfs2_schedule_truncate_log_flush(osb, 1);
 

_______________________________________________
Ocfs2-devel mailing list
Ocfs2-devel at oss.oracle.com
https://oss.oracle.com/mailman/listinfo/ocfs2-devel

You may want to add one more parameter to ocfs2_remove_btree_range indicating if the lock is already taken outside or not.
If not taken, do the lock and unlock inside; otherwise, skip lock/unlock

Will that work?

thanks,
wengang

? 2014?07?25? 13:49, Wengang ??:

Quoted 112 lines

Will it be safe to do the moving?

There are three calls of ocfs2_remove_btree_range

1 7124 fs/ocfs2/alloc.c <<ocfs2_commit_truncate>> status = ocfs2_remove_btree_range(inode, &et, trunc_cpos, 2 4479 fs/ocfs2/dir.c <<ocfs2_dx_dir_truncate>>
ret = ocfs2_remove_btree_range(dir, &et, cpos, p_cpos, clen, 0, 3 1805 fs/ocfs2/file.c <<ocfs2_remove_inode_range>> ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,

Obviously, the 2nd and the 3rd are not sub-routines of ocfs2_commit_truncate().

I think this moving is not safe.

thanks,
wengang

? 2014?07?25? 13:32, Junxiao Bi ??:

Quoted 81 lines

When running ocfs2 test suite multiple nodes reflink stress test, for a 4 nodes cluster, every unlink() for refcounted file needs about 700s.

The slow unlink is caused by the contention of refcount tree lock since all
nodes are unlink files using the same refcount tree. When the unlinking file
have many extents(over 1600 in our test), most of the extents has refcounted
flag set. In ocfs2_commit_truncate(), it will execute the following call trace
for every extents. This means it needs get and released refcount tree lock
about 1600 times. And when several nodes are do this at the same time, the
performance will be very low.
.
ocfs2_remove_btree_range() ----ocfs2_lock_refcount_tree() ------ocfs2_refcount_lock() --------__ocfs2_cluster_lock()

ocfs2_refcount_lock() is costly, move it to ocfs2_commit_truncate() to do
lock/unlock once can improve a lot performance.

Signed-off-by Junxiao Bi <[email protected]>

---
   fs/ocfs2/alloc.c |   22 ++++++++++++----------
   1 file changed, 12 insertions(+), 10 deletions(-)

fs/ocfs2/alloc.c +0 -8

index 9d8fcf2..a764363 100644
@@ -5667,13 +5667,6 @@ int ocfs2_remove_btree_range(struct inode
 *inode,
           BUG_ON(!(OCFS2_I(inode)->ip_dyn_features &
                OCFS2_HAS_REFCOUNT_FL));
   -        ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
-                           &ref_tree, NULL);
-        if (ret) {
-            mlog_errno(ret);
-            goto bail;
-        }
-
           ret = ocfs2_prepare_refcount_change_for_del(inode,
                                   refcount_loc,
                                   phys_blkno,

@@ -5755,9 +5748,6 @@ bail:
       if (meta_ac)
           ocfs2_free_alloc_context(meta_ac);
   -    if (ref_tree)
-        ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
-
       return ret;
   }
   @@ -7012,6 +7002,7 @@ int ocfs2_commit_truncate(struct

ocfs2_super *osb,

       u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
       struct ocfs2_extent_tree et;
       struct ocfs2_cached_dealloc_ctxt dealloc;
+    struct ocfs2_refcount_tree *ref_tree = NULL;
          ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
di_bh);
       ocfs2_init_dealloc_ctxt(&dealloc);

fs/ocfs2/alloc.c +8 -0

@@ -7121,6 +7112,15 @@ start:
          phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
   +    if ((flags & OCFS2_EXT_REFCOUNTED) && trunc_len && !ref_tree) {
+        status = ocfs2_lock_refcount_tree(osb, refcount_loc, 1,
+                &ref_tree, NULL);
+        if (status) {
+            mlog_errno(status);
+            goto bail;
+        }
+    }
+
       status = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
                         phys_cpos, trunc_len, flags, &dealloc,
                         refcount_loc);

fs/ocfs2/alloc.c start: · line 7143

Quoted 6 lines

fs/ocfs2/alloc.c +2 -0

@@ -7138,6 +7138,8 @@ start:
       goto start;
      bail:
+    if (ref_tree)
+        ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
          ocfs2_schedule_truncate_log_flush(osb, 1);
 

_______________________________________________
Ocfs2-devel mailing list
Ocfs2-devel at oss.oracle.com
https://oss.oracle.com/mailman/listinfo/ocfs2-devel