Re: [PATCH] use GFP_NOFS in kernel_event()

[Wu Fengguang <fengguang.wu@intel.com>]

On Thu, Apr 30, 2009 at 09:48:21PM +0800, Wu Fengguang wrote:
> On Thu, Apr 30, 2009 at 09:19:33PM +0800, Eric Paris wrote:
> > inotify: lockdep annotation when watch being removed
> > 
> > From: Eric Paris <eparis@redhat.com>
> > 
> > When a dentry is being evicted from memory pressure, if the inode associated
> > with that dentry has i_nlink == 0 we are going to drop all of the watches and
> > kick everything out.  Lockdep complains that previously holding inotify_mutex
> > we did a __GFP_FS allocation and now __GFP_FS reclaim is taking that lock.
> > There is no deadlock or danger, since we know on this code path we are
> > actually cleaning up and evicting everything.  So we move the lock into a new
> > class for clean up.
> 
> I can reproduce the bug and hence confirm that this patch works, so
> 
> Tested-by: Wu Fengguang <fengguang.wu@intel.com>

btw, I really see no point to have one GFP_KERNEL and one GFP_NOFS
sitting side by side inside kernel_event(). So this patch?

---
inotify: use consistent GFP_KERNEL in kernel_event()

kernel_event() has side by side kmem_cache_alloc(GFP_NOFS)
and kmalloc(GFP_KERNEL). Change to consistent GFP_KERNELs.

cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
---
 fs/notify/inotify/inotify_user.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

--- mm.orig/fs/notify/inotify/inotify_user.c
+++ mm/fs/notify/inotify/inotify_user.c
@@ -189,7 +189,7 @@ static struct inotify_kernel_event * ker
 {
 	struct inotify_kernel_event *kevent;
 
-	kevent = kmem_cache_alloc(event_cachep, GFP_NOFS);
+	kevent = kmem_cache_alloc(event_cachep, GFP_KERNEL);
 	if (unlikely(!kevent))
 		return NULL;