Re: [PATCH 3/9] support >32 bit ext4 filesystem block type in kernel

[Mingming Cao <cmm@us.ibm.com>]

Andrew Morton wrote:
> On Wed, 09 Aug 2006 18:21:00 -0700
> Mingming Cao <cmm@us.ibm.com> wrote:
> 
> 
>>Redefine ext4 in-kernel filesystem block type (ext4_fsblk_t) from unsigned
>>long to sector_t, to allow kernel to handle  >32 bit ext4 blocks.
>>
> 
> 
> I don't get it.

The intention is able to support 48bit ext4 on 64bit and 32bit arch 
whenevern CONFIG_LBD is enabled (or when sector_t is actually 64 bit). 
It just seems waste of memory to support 48bit in-kernel ext4 blks on 32 
bit if CONFIG_LBD is disabled, since we won't support large device anyway.

ext3_fsblk_t was a unsigned long type before, which is always 32bit on 
32bit arch. Thus we convert the ext3_fsblk_t to sector_t and depend on 
CONFIG_LBD to decide whether we need 48bit in-kernel variables.(on-disk 
block numbers are always 48 bit).

> 
> Randomly-chosen snippet:
> 
> 
>>@@ -274,7 +274,8 @@ static int find_group_orlov(struct super
>> 	freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter);
>> 	avefreei = freei / ngroups;
>> 	freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
>>-	avefreeb = freeb / ngroups;
>>+	avefreeb = freeb;
>>+	sector_div(avefreeb, ngroups);
>> 	ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter);
>> 
>> 	if ((parent == sb->s_root->d_inode) ||
> 
> 
> Here, `avefreeb' is still a 32-bit type.  Why feed it into sector_div()?
> 
avefreeb is ext3_fsblk_t type(sector_t type), which is 64bit on 64bit 
arch and 64bit on 32bit arch if CONFIG_LBD is enabled.

> 
>>@@ -303,13 +304,15 @@ static int find_group_orlov(struct super
>> 		goto fallback;
>> 	}
>> 
>>-	blocks_per_dir = (le32_to_cpu(es->s_blocks_count) - freeb) / ndirs;
>>+	blocks_per_dir = le32_to_cpu(es->s_blocks_count) - freeb;
>>+	sector_div(blocks_per_dir, ndirs);
> 
> 
> And here le32_to_cpu() is very much a 32-bit type.  Why sector_div()?

ah, sorry about the confusing, s_blocks_count (the number of blocks on 
this fs, on-disk) should be always 48bit value, and should not use 
le32_to_cpu() to read-in. That is being fixed in [patch 9/9] 64bit 
metadata support,  where we convert 32 bit s_blocks_count to sector_t, 
and also use mixro EXT4_BLOCKS_COUN(es) to read in the blocks count:

@@ -304,7 +306,7 @@ static int find_group_orlov(struct super
  		goto fallback;
  	}

-	blocks_per_dir = le32_to_cpu(es->s_blocks_count) - freeb;
+	blocks_per_dir = EXT4_BLOCKS_COUNT(es) - freeb;

This patch[3/9] is just fix the in-kernel blocks type...We trying to do 
this incrementally: first, support in-kernel blocks type >32bit, then 
change the on-disk blocks type to >32bit.

> 
> And I agree with me: we want to get all the sector_t's out of this
> filesystem.  unsigned long long, do_div()?
> 

Thanks,Mingming