From mboxrd@z Thu Jan  1 00:00:00 1970
From: James Simmons <jsimmons@infradead.org>
Date: Sun, 24 Feb 2019 16:35:23 +0000 (GMT)
Subject: [lustre-devel] [PATCH 02/37] lustre: llite: don't use
 class_setup_tunables()
In-Reply-To: <155053494490.24125.1514109985903155907.stgit@noble.brown>
References: <155053473693.24125.6976971762921761309.stgit@noble.brown>
 <155053494490.24125.1514109985903155907.stgit@noble.brown>
Message-ID: <alpine.LFD.2.21.1902241533350.15750@casper.infradead.org>
List-Id: <lustre-devel-lustre.org>
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit
To: lustre-devel@lists.lustre.org

y
> llite_kobj does not benefit from being in the
> lustre_kset kset (though it does need lustre_kset
> as a parent)

Nak.

Yes llite does benefit from being in the kset. On large clusters you
can end up with 1000s of clients so keeping the sysfs files setting 
in sync needs to be done in mass. The way Lustre does this is by 
running on the MGS server 'lctl conf_param $FSNAME.llite.lazystats=1'
or with the newer 'lctl set_param -P llite.$FSNAME-*.lazystats=1' as
an example. When this is done on the MGS that new value is first
persistently stored and then transmitted to the target nodes.
  
The magic behind this is with class_process_config(). In the code
path LCFG_PARAM, which handles the conf_param case' the main function
is class_modify_config() which is used by both llite and the obd
devices. This function first attempts to find the sysfs attribute
from the kobject and call lustre_attr_store() directly to set it.
If it can't find the attribute, which means its a debugfs file
most likely, then a uevent is created and sent off. The udev rule
then exexcutes some application to handle the debugfs settings.
For uevents to work the kobject (llite) must belong to a kset.

With LCFG_SET_PARAM, which handles the 'set_param -P' case the
master function is process_param2_config(). For this case we
always send attribute changes with uevents. Besides the requirement
of the kobject belonging to the kset this function uses 
kset_find_object() directly which if llite is not in the kset
will never be found.

> It also does not need the class_ktype type, as dynamic_kobj_ktype
> is sufficient.
> 
> So use the simple kobject_create_and_add() to initialize it.
> 
> This provides flexibility for making changes to
> class_setup_tunables().
> 
> Signed-off-by: NeilBrown <neilb@suse.com>
> ---
>  drivers/staging/lustre/lustre/llite/lproc_llite.c |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/staging/lustre/lustre/llite/lproc_llite.c b/drivers/staging/lustre/lustre/llite/lproc_llite.c
> index 8215296dc15d..78ec0fa65c58 100644
> --- a/drivers/staging/lustre/lustre/llite/lproc_llite.c
> +++ b/drivers/staging/lustre/lustre/llite/lproc_llite.c
> @@ -46,7 +46,7 @@ int llite_tunables_register(void)
>  {
>  	int rc = 0;
>  
> -	llite_kobj = class_setup_tunables("llite");
> +	llite_kobj = kobject_create_and_add("llite", &lustre_kset->kobj);
>  	if (IS_ERR(llite_kobj))
>  		return PTR_ERR(llite_kobj);
>  

        llite_kobj->kset = kset_get(&lustre_kset);
> 

In llite_tunables_unregister() the call for kobject_put()
should really be replaced with kobject_del(). This is a bug
in my original work.