From: Andreas Dilger <adilger@sun.com>
Subject: Re: PATCH: Making mb_history length a dynamic tunable
Date: Sun, 19 Apr 2009 20:22:52 -0700
Message-ID: <20090420032252.GA3209@webber.adilger.int>
References: <6601abe90904071020gdce65d2madc6df30c182c5cd@mail.gmail.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7BIT
Cc: ext4 development <linux-ext4@vger.kernel.org>
To: Curt Wohlgemuth <curtw@google.com>
Content-disposition: inline
In-reply-to: <6601abe90904071020gdce65d2madc6df30c182c5cd@mail.gmail.com>
Sender: linux-ext4-owner@vger.kernel.org

On Apr 07, 2009  10:20 -0700, Curt Wohlgemuth wrote:
> Since we frequently run in memory-constrained systems with many partitions,
> the ~68K for each partition for the mb_history buffer can be excessive.  The
> following creates a new proc file under /proc/fs/ext4/ to control the number
> of entries at mount time.

Sorry for the delay, it's been a hectic 2 weeks.  I'm totally OK with this
patch.

> If the notion of a history length tunable is okay, but the location should
> be under /sys/fs/ext4/ instead of /proc/fs/ext4/, I can change this.  The
> leftover files under /proc/fs/ext4/<partition>/ are a bit confusing to me.

Complex /proc files like this cannot be moved into /sys because it
does not support the seqfile interface for having output span more
than one page.

> ---
> --- ext4/fs/ext4/mballoc.c.orig	2009-04-07 09:13:12.000000000 -0700
> +++ ext4/fs/ext4/mballoc.c	2009-04-07 10:08:05.000000000 -0700
> @@ -334,6 +334,10 @@
>  static struct kmem_cache *ext4_pspace_cachep;
>  static struct kmem_cache *ext4_ac_cachep;
>  static struct kmem_cache *ext4_free_ext_cachep;
> +
> +#define DEFAULT_HIST_SIZE 1000
> +static int ext4_mbhist_size = DEFAULT_HIST_SIZE;
> +
>  static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
>  					ext4_group_t group);
>  static void ext4_mb_generate_from_freelist(struct super_block *sb,
> void *bitmap,
> @@ -2417,31 +2421,36 @@ static struct file_operations ext4_mb_se
>  static void ext4_mb_history_release(struct super_block *sb)
>  {
>  	struct ext4_sb_info *sbi = EXT4_SB(sb);
> +	int delete_mbhist = sbi->s_mb_history_max != 0;
> 
>  	if (sbi->s_proc != NULL) {
>  		remove_proc_entry("mb_groups", sbi->s_proc);
> -		remove_proc_entry("mb_history", sbi->s_proc);
> +		if (delete_mbhist)
> +			remove_proc_entry("mb_history", sbi->s_proc);
>  	}
> -	kfree(sbi->s_mb_history);
> +	if (delete_mbhist)
> +		kfree(sbi->s_mb_history);

If sbi->s_mb_history == NULL (as opposed to garbage) it is OK to
just call kfree() without the extra check.

>  static void ext4_mb_history_init(struct super_block *sb)
>  {
>  	struct ext4_sb_info *sbi = EXT4_SB(sb);
>  	int i;
> +	int create_mbhist = ext4_mbhist_size != 0;
> 
>  	if (sbi->s_proc != NULL) {
> -		proc_create_data("mb_history", S_IRUGO, sbi->s_proc,
> -				 &ext4_mb_seq_history_fops, sb);
> +		if (create_mbhist)
> +			proc_create_data("mb_history", S_IRUGO, sbi->s_proc,
> +					 &ext4_mb_seq_history_fops, sb);
>  		proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
>  				 &ext4_mb_seq_groups_fops, sb);
>  	}
> 
> -	sbi->s_mb_history_max = 1000;
> +	sbi->s_mb_history_max = ext4_mbhist_size;
>  	sbi->s_mb_history_cur = 0;
>  	spin_lock_init(&sbi->s_mb_history_lock);
>  	i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history);
> -	sbi->s_mb_history = kzalloc(i, GFP_KERNEL);
> +	sbi->s_mb_history = create_mbhist ? kzalloc(i, GFP_KERNEL) : NULL;
>  	/* if we can't allocate history, then we simple won't use it */
>  }

Since the s_mb_history buffers are allocated at mount time only,
and later writing to mb_hist_size does not affect their size,
it means that mb_hist_size must be set before the filesystems are
mounted.  However, that makes it impossible to tune the root
filesystem history size, and at best it is difficult to tune the
other filesystems because the value has to be set by an init
script after root mounts but before other filesystems mount.

Another possible option is to have a module parameter that can be
set when the ext4 module is inserted, so that it is sure to be
available for all filesystems, but this won't help if ext4 is
built into the kernel.

The final (though more complex) option is to add a per-fs
tunable that allows changing the history size at runtime for
each filesystem.

> @@ -2894,6 +2903,49 @@ static void release_blocks_on_commit(jou
>  	mb_debug("freed %u blocks in %u structures\n", count, count2);
>  }
> 
> +static ssize_t read_mb_hist_size(struct file *file, char __user *buf,
> +				size_t count, loff_t *ppos)
> +static ssize_t write_mb_hist_size(struct file *file, const char __user *buf,
> +				   size_t count, loff_t *ppos)

Please add an ext4_ prefix to these function names.

> +{
> +	/* This allows for 99999 entries, at 68 bytes each */
> +#define MAX_BUFSIZE 6
> +	char kbuf[MAX_BUFSIZE + 1];
> +	int value;
> +
> +	if (count) {
> +		if (count > MAX_BUFSIZE)
> +			return -EINVAL;
> +		if (copy_from_user(&kbuf, buf, count))
> +			return -EFAULT;
> +		kbuf[min(count, sizeof(kbuf))-1] = '\0';

This is could probably be avoided.  Simply initializing "kbuf" at
declaration time will ensure that there is a trailing NUL because
at most MAX_BUFSIZE bytes are copied into it.

	char kbuf[MAX_BUFSIZE + 1] = "";

> +		value = simple_strtol(kbuf, NULL, 0);
> +
> +		if (value < 0)
> +			return -EINVAL;
> +
> +		ext4_mbhist_size = value;
> +	}
> +	return count;
> +#undef MAX_BUFSIZE
> +}
> +
> +static struct file_operations ext4_mb_size_fops = {
> +	.read		= read_mb_hist_size,
> +	.write		= write_mb_hist_size,
> +};
> +
>  int __init init_ext4_mballoc(void)
>  {
>  	ext4_pspace_cachep =
> @@ -2912,6 +2964,9 @@ int __init init_ext4_mballoc(void)
>  		return -ENOMEM;
>  	}
> 
> +	proc_create("mb_hist_size", S_IRUGO | S_IWUGO, ext4_proc_root,
> +			 &ext4_mb_size_fops);
> +
>  	ext4_free_ext_cachep =
>  		kmem_cache_create("ext4_free_block_extents",
>  				     sizeof(struct ext4_free_data),

Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.