From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-kernel-owner@vger.kernel.org>
Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand
	id S1752392AbcBTQfb (ORCPT <rfc822;w@1wt.eu>);
	Sat, 20 Feb 2016 11:35:31 -0500
Received: from mail-ob0-f175.google.com ([209.85.214.175]:36005 "EHLO
	mail-ob0-f175.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
	with ESMTP id S1750987AbcBTQfa (ORCPT
	<rfc822;linux-kernel@vger.kernel.org>);
	Sat, 20 Feb 2016 11:35:30 -0500
Subject: Re: [RFC] sysfs: defer instantiation of default attrs
To: Edward Cree <ec429@cantab.net>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>
References: <alpine.DEB.2.20.1602201041100.23349@dev-full.penguins>
Cc: linux-kernel@vger.kernel.org
From: David Ahern <dsa@cumulusnetworks.com>
Message-ID: <56C895CD.9070702@cumulusnetworks.com>
Date: Sat, 20 Feb 2016 09:35:25 -0700
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.11; rv:38.0)
 Gecko/20100101 Thunderbird/38.6.0
MIME-Version: 1.0
In-Reply-To: <alpine.DEB.2.20.1602201041100.23349@dev-full.penguins>
Content-Type: text/plain; charset=windows-1252; format=flowed
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
List-ID: <linux-kernel.vger.kernel.org>
X-Mailing-List: linux-kernel@vger.kernel.org

Hi Ed:

Thank for taking the to look into this.

On 2/20/16 3:42 AM, Edward Cree wrote:
> Testing this in a VM, and with udevd disabled (being too lazy to do it
>   properly), created 1024 bridges.  On ls'ing
>   /sys/class/net/bridge*/queues/*/, saw free memory drop by 64kB,
>   suggesting that much had been saved by deferral.  It's not very much,
>   hopefully deferring attribute groups will do better.

That sounds consistent with what I found by cutting out attributes in 
netdev_register_kobject. e.g., from my patch:

diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c
index b6c8a6629b39..f0df828a2f20 100644
--- a/net/core/net-sysfs.c
+++ b/net/core/net-sysfs.c
@@ -1524,7 +1524,8 @@ int netdev_register_kobject(struct net_device *ndev)
         int error = 0;

         device_initialize(dev);
-       dev->class = &net_class;
+       if (!netif_is_lwt(ndev))
+               dev->class = &net_class;
         dev->platform_data = ndev;
         dev->groups = groups;

@@ -1535,7 +1536,8 @@ int netdev_register_kobject(struct net_device *ndev)
         if (*groups)
                 groups++;

-       *groups++ = &netstat_group;
+       if (!netif_is_lwt(ndev)) {
+               *groups++ = &netstat_group;

  #if IS_ENABLED(CONFIG_WIRELESS_EXT) || IS_ENABLED(CONFIG_CFG80211)
         if (ndev->ieee80211_ptr)


The big memory savings came in when I bypassed register_queue_kobjects:

         error = device_add(dev);
         if (error)
                 return error;

-       error = register_queue_kobjects(ndev);
-       if (error) {
-               device_del(dev);
-               return error;
+       if (!netif_is_lwt(ndev)) {
+               error = register_queue_kobjects(ndev);
+               if (error) {
+                       device_del(dev);
+                       return error;
+               }
         }

         pm_runtime_set_memalloc_noio(dev, true);


Yes, this is a bit Draconian and does have impacts to tools looking for 
/sys files but the return is huge when you look at 1000's of netdevices 
(ports, vlans, bridges, bonds).