GFS2: Reduce number of log flushes

Hi,

This patch reduces the number of log flushes.

Regards,

Bob Peterson
Red Hat File Systems

---

index 4bdcf37..7d795e7 100644
@@ -118,7 +118,13 @@ void gfs2_ail_flush(struct gfs2_glock *gl, bool fsync)
 		return;
 	__gfs2_ail_flush(gl, fsync);
 	gfs2_trans_end(sdp);
-	gfs2_log_flush(sdp, NULL);
+	/* if fsync is true, we were called from gfs2_fsync, which means
+	   we just called sync_inode_metadata, which calls sync_inode,
+	   which calls writeback_single_inode, which calls write_inode,
+	   which calls write_inode which calls gfs2_log_flush. So the call
+	   is redundant. */
+	if (!fsync)
+		gfs2_log_flush(sdp, NULL);
 }
 
 /**

index 3cbac68..5904ce0 100644
@@ -1542,7 +1542,11 @@ static void gfs2_evict_inode(struct inode *inode)
 	goto out_unlock;
 
 out_truncate:
-	gfs2_log_flush(sdp, ip->i_gl);
+	/* write_inode_now calls writeback_single_inode, which calls
+	   gfs2_write_inode. gfs2_write_inode calls gfs2_log_flush if
+	   (wbc->sync_mode == WB_SYNC_ALL), which it is since we pass 1.
+	   therefore, this call to gfs2_log_flush is redundant:
+	   gfs2_log_flush(sdp, ip->i_gl); */
 	write_inode_now(inode, 1);
 	gfs2_ail_flush(ip->i_gl, 0);
 

Hi,

----- Original Message -----
| What ensures that the revokes are on disk?
(snip)

| This one looks like it is probably correct, however there is no point
| in
| adding a comment to refer to code that has just been removed. Just
| remove that call and put the explanation in the patch description, or
| maybe include just a short comment so say that there is a log flush
| implicit in the write_inode_now() call,
| 
| Steve.

You're right about the revokes; my bad. Here is a respin, with just a comment in the patch rather than in the code:

Hi,

In function gfs2_evict_inode, there was a call to gfs2_log_flush that is unnecessary because it immediately calls write_inode_now which makes another call to gfs2_log_flush. This patch eliminates the redundant log flush.

Regards,

Bob Peterson
Red Hat GFS

---

index 3cbac68..ba4fde5 100644
@@ -1542,7 +1542,6 @@ static void gfs2_evict_inode(struct inode *inode)
 	goto out_unlock;
 
 out_truncate:
-	gfs2_log_flush(sdp, ip->i_gl);
 	write_inode_now(inode, 1);
 	gfs2_ail_flush(ip->i_gl, 0);
 

----- Original Message -----

| ----- Original Message -----
| | What ensures that the revokes are on disk?
| (snip)
| | This one looks like it is probably correct, however there is no
| | point
| | in
| | adding a comment to refer to code that has just been removed. Just
| | remove that call and put the explanation in the patch description,
| | or
| | maybe include just a short comment so say that there is a log flush
| | implicit in the write_inode_now() call,
| | 
| | Steve.
| 
| You're right about the revokes; my bad. Here is a respin, with just
| a comment in the patch rather than in the code:
| 
| Hi,
| 
| In function gfs2_evict_inode, there was a call to gfs2_log_flush
| that is unnecessary because it immediately calls write_inode_now
| which makes another call to gfs2_log_flush. This patch eliminates
| the redundant log flush.
| 
| Regards,
| 
| Bob Peterson
| Red Hat GFS
| 
| Signed-off-by: Bob Peterson <[email protected]>
| ---
| diff --git a/fs/gfs2/super.c b/fs/gfs2/super.c
| index 3cbac68..ba4fde5 100644
| --- a/fs/gfs2/super.c
| +++ b/fs/gfs2/super.c
| @@ -1542,7 +1542,6 @@ static void gfs2_evict_inode(struct inode
| *inode)
|  	goto out_unlock;
|  
|  out_truncate:
| -	gfs2_log_flush(sdp, ip->i_gl);
|  	write_inode_now(inode, 1);
|  	gfs2_ail_flush(ip->i_gl, 0);
|  
| 

Hi,

I've still not heard back on this revised patch from 09 Aug, and it has not gone into -nmw. Perhaps it was missed?

Regards,

Bob Peterson
Red Hat File Systems

Hi,

Yes, I think that might have fallen off the stack. Apologies for the delay. However I think we should check with Ben as this is the same bit of code that he is looking at currently,

Steve.