Subject: RE: 2.6.10-mm3 scaling problem with inotify
From: John McCutchan <ttb@tentacle.dhs.org>
To: "Zou, Nanhai" <nanhai.zou@intel.com>
Cc: linux-kernel@vger.kernel.org, Robert Love <rml@novell.com>,
       hawkes@tomahawk.engr.sgi.com
In-Reply-To: <894E37DECA393E4D9374E0ACBBE7427013CA7E@pdsmsx402.ccr.corp.intel.com>
References: <894E37DECA393E4D9374E0ACBBE7427013CA7E@pdsmsx402.ccr.corp.intel.com>
Content-Type: text/plain
Content-Transfer-Encoding: 7bit
Date: Thu, 13 Jan 2005 21:36:20 -0500
Message-Id: <1105670180.25359.2.camel@vertex>
Mime-Version: 1.0
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1201
Lines: 32

On Fri, 2005-01-14 at 10:31 +0800, Zou, Nanhai wrote:
>  
>  There is still a little difference between your implement in
>  inotify_dentry_parent_queue_event from dnotify_parent
>  
>  In dnotify_parent, if parent is not watching the event, the code will
> not fall
>  through dget and dput path.
>  
>  While in inotify_dentry_parent_queue_event kernel will go dget and dput
> even
>  if (inode->inotify_data == NULL).
>  
>  While dget and dput will introduce a lot of atomic operations..
>  And the most important, dput will grab global dcache_lock...,
>  I think that is the reason why John Hawkes saw great performance drop.
>  
>  Simply follow dnotify_parent, only call dget and dput when
> inode->inotify_data != NULL will solve this problem I think.
> 

Yeah, the old code was written before inode->inotify_data existed.
Robert caught this before he sent his patch. His patch should fix this
regression.

-- 
John McCutchan <ttb@tentacle.dhs.org>
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/