From: Andreas Dilger <adilger@sun.com>
Subject: Re: [PATCH] UPDATED: types fixup for mballoc
Date: Thu, 3 Jan 2008 12:17:33 -0700
Message-ID: <20080103191733.GJ3351@webber.adilger.int>
References: <473CCEB8.5060201@redhat.com> <473D2147.9030504@redhat.com> <477BEDAC.1030802@redhat.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Cc: ext4 development <linux-ext4@vger.kernel.org>,
	Alex Tomas <alex@clusterfs.com>
To: Eric Sandeen <sandeen@redhat.com>
Return-path: <linux-ext4-owner@vger.kernel.org>
Received: from mail.clusterfs.com ([74.0.229.162]:37845 "EHLO
	mail.clusterfs.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
	with ESMTP id S1751784AbYACTRi (ORCPT
	<rfc822;linux-ext4@vger.kernel.org>); Thu, 3 Jan 2008 14:17:38 -0500
Content-Disposition: inline
In-Reply-To: <477BEDAC.1030802@redhat.com>
Sender: linux-ext4-owner@vger.kernel.org
List-ID: <linux-ext4.vger.kernel.org>

On Jan 02, 2008  14:01 -0600, Eric Sandeen wrote:
> This patch:
> 
> changes fe_len to an int, I don't think we need it to be a long,
> looking at how it's used (should it be a grpblk_t?)  Also change
> anything assigned to return value of mb_find_extent, since it returns
> fe_len.
> 
> changes anything that does groupno * EXT4_BLOCKS_PER_GROUP
> or pa->pa_pstart + <whatever> to an ext4_fsblk_t
> 
> fixes up any related formats
> 
> The change to ext4_mb_scan_aligned to avoid a 64-bit modulo
> could use an extra look I think.
>
> @@ -1732,17 +1735,18 @@ static void ext4_mb_scan_aligned(struct 
>  	BUG_ON(sbi->s_stripe == 0);
>  
> -	/* find first stripe-aligned block */
> -	i = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
> -		+ le32_to_cpu(sbi->s_es->s_first_data_block);
> -	i = ((i + sbi->s_stripe - 1) / sbi->s_stripe) * sbi->s_stripe;
> -	i = (i - le32_to_cpu(sbi->s_es->s_first_data_block))
> -			% EXT4_BLOCKS_PER_GROUP(sb);

> +	/* find first stripe-aligned block in group */
> +	a = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb)
> +		+ le32_to_cpu(sbi->s_es->s_first_data_block);
> +	a = (a + sbi->s_stripe - 1) & ~((ext4_fsblk_t)sbi->s_stripe - 1);
> +	a = a - le32_to_cpu(sbi->s_es->s_first_data_block);
> +	i = do_div(a, EXT4_BLOCKS_PER_GROUP(sb));

I don't think this is correct...  This code should only be used if s_stripe
is NOT a power-of-two value, otherwise we can just use the buddy maps to
find aligned chunks.  As a result I don't think that "& (s_stripe - 1)"
change is equivalent to "/ s_stripe * s_stripe".

>  	while (i < sb->s_blocksize * 8) {
>  		if (!mb_test_bit(i, bitmap)) {

Hrmmm, I thought this should be "while (i < EXT4_BLOCKS_PER_GROUP(sb))"
and not "sb->s_blocksize * 8"?  Did that fix get lost somewhere?

There are a few other changes in the patch related to this fix:
https://bugzilla.lustre.org/attachment.cgi?id=13275

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