Patchwork e2fsprogs: allocate inode table wholly within group

login
register
mail settings
Submitter Eric Sandeen
Date July 2, 2013, 7:14 p.m.
Message ID <51D3269B.5080608@redhat.com>
Download mbox | patch
Permalink /patch/256476/
State New
Headers show

Comments

Eric Sandeen - July 2, 2013, 7:14 p.m.
Building e2fsprogs 1.42.8 on ppc, I got this:

r_1024_small_bg: ext2 1024 blocksize with small block groups: failed

Because during the resize step it did this:

Itable move group 1 block 1030->1092 (diff 62)

but during e2fsck it found:

/tmp/e2fsprogs-tmp.uiFhgP: Inode table for group 1 is not in group.  (block 1092)

i.e. from dumpe2fs we can see:

Group 1: (Blocks 1025-1110)
  Backup superblock at 1025, Group descriptors at 1026-1026
  Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
  Inode table at 1092-1123 (+67)
                      ^^^^ beyond end of block group

ext2fs_allocate_group_table() currently sends the last block of the
group as an acceptable *starting* point for the inode table allocation.

Because the inode table may be several blocks, and must reside wholly
within the group, move the last acceptable starting block back by this
amount so that the allocated range cannot extend past the end of the
group.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
---

p.s. I *think* this fixes it; I've not convinced myself that the failure
is 100% deterministic, so extra eyeballs on the patch are welcome.

Thanks,
-Eric


--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Eric Sandeen - July 4, 2013, 2:19 p.m.
On 7/2/13 3:14 PM, Eric Sandeen wrote:
> Building e2fsprogs 1.42.8 on ppc, I got this:
> 
> r_1024_small_bg: ext2 1024 blocksize with small block groups: failed

You can add a RH FTBFS bug to the commit now if you like, 

[Bug 980519] New: FTBFS: self checks failure: Tests failed: r_1024_small_bg

https://bugzilla.redhat.com/show_bug.cgi?id=980519

since Karsten Hopp has now reported it for Rawhide as well.

-Eric


> Because during the resize step it did this:
> 
> Itable move group 1 block 1030->1092 (diff 62)
> 
> but during e2fsck it found:
> 
> /tmp/e2fsprogs-tmp.uiFhgP: Inode table for group 1 is not in group.  (block 1092)
> 
> i.e. from dumpe2fs we can see:
> 
> Group 1: (Blocks 1025-1110)
>   Backup superblock at 1025, Group descriptors at 1026-1026
>   Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
>   Inode table at 1092-1123 (+67)
>                       ^^^^ beyond end of block group
> 
> ext2fs_allocate_group_table() currently sends the last block of the
> group as an acceptable *starting* point for the inode table allocation.
> 
> Because the inode table may be several blocks, and must reside wholly
> within the group, move the last acceptable starting block back by this
> amount so that the allocated range cannot extend past the end of the
> group.
> 
> Signed-off-by: Eric Sandeen <sandeen@redhat.com>
> ---
> 
> p.s. I *think* this fixes it; I've not convinced myself that the failure
> is 100% deterministic, so extra eyeballs on the patch are welcome.
> 
> Thanks,
> -Eric
> 
> Index: e2fsprogs-1.42.8/lib/ext2fs/alloc_tables.c
> ===================================================================
> --- e2fsprogs-1.42.8.orig/lib/ext2fs/alloc_tables.c
> +++ e2fsprogs-1.42.8/lib/ext2fs/alloc_tables.c
> @@ -200,6 +200,9 @@ errcode_t ext2fs_allocate_group_table(ex
>  	}
>  
>  	if (!ext2fs_inode_table_loc(fs, group)) {
> +		/* Inode table must fit within group, so adjust last blk */
> +		last_blk -= (fs->inode_blocks_per_group - 1);
> +
>  		retval = ext2fs_get_free_blocks2(fs, group_blk, last_blk,
>  						fs->inode_blocks_per_group,
>  						bmap, &new_blk);
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Theodore Ts'o - July 7, 2013, 3:53 p.m.
On Tue, Jul 02, 2013 at 03:14:35PM -0400, Eric Sandeen wrote:
> Building e2fsprogs 1.42.8 on ppc, I got this:
> 
> r_1024_small_bg: ext2 1024 blocksize with small block groups: failed
> 
> Because during the resize step it did this:
> 
> Itable move group 1 block 1030->1092 (diff 62)
> 
> but during e2fsck it found:
> 
> /tmp/e2fsprogs-tmp.uiFhgP: Inode table for group 1 is not in group.  (block 1092)
> 
> i.e. from dumpe2fs we can see:
> 
> Group 1: (Blocks 1025-1110)
>   Backup superblock at 1025, Group descriptors at 1026-1026
>   Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
>   Inode table at 1092-1123 (+67)
>                       ^^^^ beyond end of block group

There seems to be something wrong here.  The test file system was
created like this:

  mke2fs -t ext2 -O ^resize_inode -b 1024 -g 1024 -qF /tmp/foo.img 64M

The file system hence should have 64 block groups, and dumpe2fs before
the resize looks like this on an x86 system:

Group 1: (Blocks 1025-2048)
  Backup superblock at 1025, Group descriptors at 1026-1027
  Block bitmap at 1028 (+3), Inode bitmap at 1029 (+4)
  Inode table at 1030-1061 (+5)

... and after:

Group 1: (Blocks 1025-2048)
  Backup superblock at 1025, Group descriptors at 1026-1089
  Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
  Inode table at 1092-1123 (+67)

Note the range of block group #1: 1025-2048, whereas on the PPC,
apparently the range is quite different: Group 1: (Blocks 1025-1110)

So there's something else going really wrong here....

   	   	     	  	       	     - Ted
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Theodore Ts'o - July 7, 2013, 11:34 p.m.
On Sun, Jul 07, 2013 at 11:53:16AM -0400, Theodore Ts'o wrote:
> There seems to be something wrong here.  The test file system was
> created like this:
> 
>   mke2fs -t ext2 -O ^resize_inode -b 1024 -g 1024 -qF /tmp/foo.img 64M
> 
> The file system hence should have 64 block groups, and dumpe2fs before
> the resize looks like this on an x86 system:
> 
> Group 1: (Blocks 1025-2048)
>   Backup superblock at 1025, Group descriptors at 1026-1027
>   Block bitmap at 1028 (+3), Inode bitmap at 1029 (+4)
>   Inode table at 1030-1061 (+5)
> 
> ... and after:
> 
> Group 1: (Blocks 1025-2048)
>   Backup superblock at 1025, Group descriptors at 1026-1089
>   Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
>   Inode table at 1092-1123 (+67)
> 
> Note the range of block group #1: 1025-2048, whereas on the PPC,
> apparently the range is quite different: Group 1: (Blocks 1025-1110)
> 
> So there's something else going really wrong here....

I just tried building e2fsprogs 1.42.8 on a powerpc system in a Debian
unstable (sid) chroot, and I'm not able to reproduce this test
failure.  I tried building using both gcc 4.6.4 and gcc 4.8, in case
it was a compiler bug.

So again, there's something really, REALLY wrong....

							- Ted
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Eric Sandeen - July 8, 2013, 1:59 a.m.
On Jul 7, 2013, at 10:53 AM, "Theodore Ts'o" <tytso@mit.edu> wrote:

> On Tue, Jul 02, 2013 at 03:14:35PM -0400, Eric Sandeen wrote:
>> Building e2fsprogs 1.42.8 on ppc, I got this:
>> 
>> r_1024_small_bg: ext2 1024 blocksize with small block groups: failed
>> 
>> Because during the resize step it did this:
>> 
>> Itable move group 1 block 1030->1092 (diff 62)
>> 
>> but during e2fsck it found:
>> 
>> /tmp/e2fsprogs-tmp.uiFhgP: Inode table for group 1 is not in group.  (block 1092)
>> 
>> i.e. from dumpe2fs we can see:
>> 
>> Group 1: (Blocks 1025-1110)
>>  Backup superblock at 1025, Group descriptors at 1026-1026
>>  Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
>>  Inode table at 1092-1123 (+67)
>>                      ^^^^ beyond end of block group
> 
> There seems to be something wrong here.  The test file system was
> created like this:
> 
>  mke2fs -t ext2 -O ^resize_inode -b 1024 -g 1024 -qF /tmp/foo.img 64M
> 
> The file system hence should have 64 block groups, and dumpe2fs before
> the resize looks like this on an x86 system:
> 
> Group 1: (Blocks 1025-2048)
>  Backup superblock at 1025, Group descriptors at 1026-1027
>  Block bitmap at 1028 (+3), Inode bitmap at 1029 (+4)
>  Inode table at 1030-1061 (+5)
> 
> ... and after:
> 
> Group 1: (Blocks 1025-2048)
>  Backup superblock at 1025, Group descriptors at 1026-1089
>  Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
>  Inode table at 1092-1123 (+67)
> 
> Note the range of block group #1: 1025-2048, whereas on the PPC,
> apparently the range is quite different: Group 1: (Blocks 1025-1110)
> 
> So there's something else going really wrong here....
> 
>                                             - Ted

Perhaps so, I'll look into it more.  But nonetheless isn't the patch correct?

Don't we need to ensure the inode table is wholly within be group, however we may have gotten here?

-Eric--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Eric Sandeen - July 8, 2013, 6:34 p.m.
On 7/7/13 10:53 AM, Theodore Ts'o wrote:
> On Tue, Jul 02, 2013 at 03:14:35PM -0400, Eric Sandeen wrote:
>> Building e2fsprogs 1.42.8 on ppc, I got this:
>>
>> r_1024_small_bg: ext2 1024 blocksize with small block groups: failed
>>
>> Because during the resize step it did this:
>>
>> Itable move group 1 block 1030->1092 (diff 62)
>>
>> but during e2fsck it found:
>>
>> /tmp/e2fsprogs-tmp.uiFhgP: Inode table for group 1 is not in group.  (block 1092)
>>
>> i.e. from dumpe2fs we can see:
>>
>> Group 1: (Blocks 1025-1110)
>>   Backup superblock at 1025, Group descriptors at 1026-1026
>>   Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
>>   Inode table at 1092-1123 (+67)
>>                       ^^^^ beyond end of block group
> 
> There seems to be something wrong here.  The test file system was
> created like this:
> 
>   mke2fs -t ext2 -O ^resize_inode -b 1024 -g 1024 -qF /tmp/foo.img 64M
> 
> The file system hence should have 64 block groups, and dumpe2fs before
> the resize looks like this on an x86 system:
> 
> Group 1: (Blocks 1025-2048)
>   Backup superblock at 1025, Group descriptors at 1026-1027
>   Block bitmap at 1028 (+3), Inode bitmap at 1029 (+4)
>   Inode table at 1030-1061 (+5)
> 
> ... and after:
> 
> Group 1: (Blocks 1025-2048)
>   Backup superblock at 1025, Group descriptors at 1026-1089
>   Block bitmap at 1090 (+65), Inode bitmap at 1091 (+66)
>   Inode table at 1092-1123 (+67)
> 
> Note the range of block group #1: 1025-2048, whereas on the PPC,
> apparently the range is quite different: Group 1: (Blocks 1025-1110)
> 
> So there's something else going really wrong here....

I don't think so - that's only after the shrink, right?

However, I understand your confusion.  The test is not at all obvious
about what it's doing:

> test_description="ext2 1024 blocksize with small block groups"
> FEATURES="-t ext2 -O ^resize_inode -b 1024 -g 1024"
> SIZE_1=64M
> SIZE_2=2G
> LOG=$test_name.log
> E2FSCK=../e2fsck/e2fsck
> 
> 
> . $cmd_dir/scripts/resize_test

Um, ok, so we "resize test" with those parameters.

Looking at that script, which sadly has no comments, it invokes
a resize to $SIZE_2 (2G) first, but then does something different:

# grep RESIZE tests/scripts/resize_test | grep -v echo
if ! $RESIZE2FS $RESIZE2FS_OPTS -d 31 $TMPFILE $SIZE_2 >> $LOG 2>&1
if ! $RESIZE2FS $RESIZE2FS_OPTS -d 31 -M $TMPFILE $SIZE_2 >> $LOG 2>&1
if ! $RESIZE2FS $RESIZE2FS_OPTS -d 31 -M $TMPFILE $SIZE_2 >> $LOG 2>&1
if ! $RESIZE2FS $RESIZE2FS_OPTS -d 31 -M $TMPFILE $SIZE_2 >> $LOG 2>&1

so it does 4 resizes; the last 3 are using "-M" which says "go to the
minimum size."   Which is a weird thing to do when also specifying
an actual size:

# grep "\-M" tests/r_1024_small_bg.log 
../resize/resize2fs -d 31 -M /tmp/e2fsprogs-tmp.LoNby5 2G
../resize/resize2fs -d 31 -M /tmp/e2fsprogs-tmp.LoNby5 2G
../resize/resize2fs -d 31 -M /tmp/e2fsprogs-tmp.LoNby5 2G

But the 2G is ignored.... so it's re-resizing to minimum sizes.  On x86:

# grep -w "resize\|is now" tests/r_1024_small_bg.log 
../resize/resize2fs -d 31 /tmp/e2fsprogs-tmp.LoNby5 2G
The filesystem on /tmp/e2fsprogs-tmp.LoNby5 is now 2097152 blocks long.
../resize/resize2fs -d 31 -M /tmp/e2fsprogs-tmp.LoNby5 2G
The filesystem on /tmp/e2fsprogs-tmp.LoNby5 is now 1174 blocks long.
../resize/resize2fs -d 31 -M /tmp/e2fsprogs-tmp.LoNby5 2G
The filesystem on /tmp/e2fsprogs-tmp.LoNby5 is now 1002 blocks long.

on the failing PPC build:

# grep -w "resize\|is now" tests/r_1024_small_bg.log 
../resize/resize2fs -d 31 /tmp/e2fsprogs-tmp.4KTO7a 2G
The filesystem on /tmp/e2fsprogs-tmp.4KTO7a is now 2097152 blocks long.
../resize/resize2fs -d 31 -M /tmp/e2fsprogs-tmp.4KTO7a 2G
The filesystem on /tmp/e2fsprogs-tmp.4KTO7a is now 1247 blocks long.
../resize/resize2fs -d 31 -M /tmp/e2fsprogs-tmp.4KTO7a 2G
The filesystem on /tmp/e2fsprogs-tmp.4KTO7a is now 1121 blocks long.

so they're both shrinking quite a bit.  But I think due to the vagaries
of the initial filesystem population, the test is behaving differently
and just happening to hit this corner case.

On the last failing run, the minimum size happens to be calculated
at 1121 blocks; with debug printfs:

> ../resize/resize2fs -d 31 -M /tmp/e2fsprogs-tmp.gB4G7E 2G
> resize2fs 1.42.8 (20-Jun-2013)
> forcing to min size 1121
> Resizing the filesystem on /tmp/e2fsprogs-tmp.gB4G7E to 1121 (1k) blocks.

But what's weird is that in this resize run it didn't actually
move the inode table; that was done in a prior minimization run,
and in this case it seems to have simply shrunk past the existing
inode table end.  Argh.  Still looking.

-Eric

--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

Index: e2fsprogs-1.42.8/lib/ext2fs/alloc_tables.c
===================================================================
--- e2fsprogs-1.42.8.orig/lib/ext2fs/alloc_tables.c
+++ e2fsprogs-1.42.8/lib/ext2fs/alloc_tables.c
@@ -200,6 +200,9 @@  errcode_t ext2fs_allocate_group_table(ex
 	}
 
 	if (!ext2fs_inode_table_loc(fs, group)) {
+		/* Inode table must fit within group, so adjust last blk */
+		last_blk -= (fs->inode_blocks_per_group - 1);
+
 		retval = ext2fs_get_free_blocks2(fs, group_blk, last_blk,
 						fs->inode_blocks_per_group,
 						bmap, &new_blk);