Re: [PATCH 4/6] writeback: fix writeback_inodes_wb from writeback_inodes_sb

[Christoph Hellwig <hch@lst.de>]

On Tue, Jun 08, 2010 at 09:51:39PM +0200, Jens Axboe wrote:
> On 08/06/10 18.14, Christoph Hellwig wrote:
> > When we call writeback_inodes_wb from writeback_inodes_sb we always have
> > s_umount held, which currently makes the whole operation a no-op.
> > 
> > But if we are called to write out inodes for a specific superblock we always
> > have s_umount held, so replace the incorrect logic checking for WB_SYNC_ALL
> > which only worked by coincidence with the proper check for an explicit
> > superblock argument.
> 
> This is tons better than the pinning, I like it a lot.

Unfortunately I accidentally removed the requeue_io call when we fail
to pin the inode.  This leads to softlockups after heavy I/O load.

Please fold the patch below into this one, or if not possible add it
to the end of the series.

---
From: Christoph Hellwig <hch@lst.de>
Subject: [PATCH] writeback: add missing requeue_io in writeback_inodes_wb

In "writeback: fix writeback_inodes_wb from writeback_inodes_sb" I
accidentally removed the requeue_io if we need to skip a superblock
because we can't pin it.  Add it back, otherwise we're getting spurious
lockups after multiple xfstests runs.

Signed-off-by: Christoph Hellwig <hch@lst.de>

Index: linux-2.6/fs/fs-writeback.c
===================================================================
--- linux-2.6.orig/fs/fs-writeback.c	2010-06-09 12:09:42.462024800 +0200
+++ linux-2.6/fs/fs-writeback.c	2010-06-09 12:09:47.830005523 +0200
@@ -643,8 +643,10 @@ static void writeback_inodes_wb(struct b
 
 			ret = writeback_sb_inodes(sb, wb, wbc);
 		} else {
-			if (!pin_sb_for_writeback(sb))
+			if (!pin_sb_for_writeback(sb)) {
+				requeue_io(inode);
 				continue;
+			}
 			ret = writeback_sb_inodes(sb, wb, wbc);
 			drop_super(sb);
 		}