Date: Sat, 22 Oct 2016 10:00:07 +1100
From: Dave Chinner <david@fromorbit.com>
To: Shaohua Li <shli@fb.com>
Cc: linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
        Kernel-team@fb.com, viro@zeniv.linux.org.uk
Subject: Re: [RFC] put more pressure on proc/sysfs slab shrink
Message-ID: <20161021230007.GV23194@dastard>
References: <bca13f1ae6a72f0d126cd7e9ede11baaa2b81064.1477081587.git.shli@fb.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <bca13f1ae6a72f0d126cd7e9ede11baaa2b81064.1477081587.git.shli@fb.com>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1730
Lines: 48

On Fri, Oct 21, 2016 at 01:35:14PM -0700, Shaohua Li wrote:
> In our systems, proc/sysfs inode/dentry cache use more than 1G memory
> even memory pressure is high sometimes. Since proc/sysfs is in-memory
> filesystem, rebuilding the cache is fast. There is no point proc/sysfs
> and disk fs have equal pressure for slab shrink.
> 
> One idea is directly discarding proc/sysfs inode/dentry cache rightly
> after the proc/sysfs file is closed. But the discarding will make
> proc/sysfs file open slower next time, which is 20x slower in my test if
> multiple applications are accessing proc files. This patch doesn't go
> that far. Instead, just put more pressure to shrink proc/sysfs slabs.
> 
> Signed-off-by: Shaohua Li <shli@fb.com>
> ---
>  fs/kernfs/mount.c | 2 ++
>  fs/proc/inode.c   | 2 ++
>  2 files changed, 4 insertions(+)
> 
> diff --git a/fs/kernfs/mount.c b/fs/kernfs/mount.c
> index d5b149a..5b4e747 100644
> --- a/fs/kernfs/mount.c
> +++ b/fs/kernfs/mount.c
> @@ -161,6 +161,8 @@ static int kernfs_fill_super(struct super_block *sb, unsigned long magic)
>  	sb->s_xattr = kernfs_xattr_handlers;
>  	sb->s_time_gran = 1;
>  
> +	sb->s_shrink.seeks = 1;
> +	sb->s_shrink.batch = 0;

This sort of thing needs comments as to why they are being changed.
Otherwise the next person who comes along to do shrinker
modifications won't have a clue about why this magic exists.

Also, I don't think s_shrink.batch = 0 does what you think it does.
The superblock batch size default of 1024 is more efficient than
setting sb->s_shrink.batch = 0 as that makes the shrinker use
SHRINK_BATCH:

#define SHRINK_BATCH 128

i.e. it does less work per batch so has more overhead....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com