Extended partition mapping wrong size

Extended partition mapping wrong size

[Phillip Susi]

I've been investigating a problem I ran into trying to create partitions
in sector mode and found that the first logical partition can not begin
on the very first sector of the extended partition, immediately
following the EBR.  This apparently is because the kernel creates a dev
node to represent the extended partition and sizes it to two sectors.  I
could have sworn that the kernel did not used to create a device for the
extended partition itself, but I wondered why it was 2 sectors long.  I
found this:

from fs/partitions/msdos.c:

            /* prevent someone doing mkfs or mkswap on an
               extended partition, but leave room for LILO */
            put_partition(state, slot, start, size == 1 ? 1 : 2);

This appears to set the size of the device to 2 sectors, unless the
extended partition is only 1 sector long.  Shouldn't the size be
whatever length there is between the start of the extended partition,
and the first logical partition it contains?  So if there are 63 sectors
there, as is the usual case when using cylinder alignment, then the
device node would expose all of those, but if there is no space, then
the device node should only be 1 sector for the EBR, otherwise it
overlays the first sector of the logical partition.

Re: Extended partition mapping wrong size

[Karel Zak]

On Mon, Mar 29, 2010 at 12:11:23PM -0400, Phillip Susi wrote:
> I've been investigating a problem I ran into trying to create partitions
> in sector mode and found that the first logical partition can not begin
> on the very first sector of the extended partition, immediately
> following the EBR.  This apparently is because the kernel creates a dev
> node to represent the extended partition and sizes it to two sectors.

This is probably kernel bug. It's really insane that the extended
pseudo partition overflows to the next logical partition.

> could have sworn that the kernel did not used to create a device for the
> extended partition itself, but I wondered why it was 2 sectors long.  I
> found this:
> 
> from fs/partitions/msdos.c:
> 
>             /* prevent someone doing mkfs or mkswap on an
>                extended partition, but leave room for LILO */
>             put_partition(state, slot, start, size == 1 ? 1 : 2);
>
> This appears to set the size of the device to 2 sectors, unless the
> extended partition is only 1 sector long.  Shouldn't the size be
> whatever length there is between the start of the extended partition,
> and the first logical partition it contains?  So if there are 63 sectors
> there, as is the usual case when using cylinder alignment, then the

 Please no. I think the size should not be more than 2 sectors (1024
 bytes). The current concept works for years and we have in userspace 
 /etc/partitions parsers that use "if (blocks <= 1)" to detect
 extended partitions.

 The other problem are mkfs programs, the space used for alignment
 could be 1MiB (or more) -- it's enough many mkfs programs.


 BTW, Linux does not use this policy for the others nested partition
 tables (e.g. Solaris, BSD, Minix, ...). The extended dos partition
 table is exception. The primary partitions for the others nested PT
 are exported to the system with its real size :-)

    Karel

-- 
 Karel Zak  <kzak@redhat.com>
 http://karelzak.blogspot.com

Re: Extended partition mapping wrong size

[Phillip Susi]

On 4/6/2010 7:47 AM, Karel Zak wrote:
> This is probably kernel bug. It's really insane that the extended
> pseudo partition overflows to the next logical partition.

Indeed.

>  Please no. I think the size should not be more than 2 sectors (1024
>  bytes). The current concept works for years and we have in userspace 
>  /etc/partitions parsers that use "if (blocks <= 1)" to detect
>  extended partitions.

Could you elaborate a bit on this?  What programs have such tests and
what would they do differently if it were larger?

>  The other problem are mkfs programs, the space used for alignment
>  could be 1MiB (or more) -- it's enough many mkfs programs.

What's wrong with that?  If you REALLY want to, there's no reason you
can't create a tiny fs there.  Then again, I could swear that once upon
a time the kernel simply did not bother creating a dev node for the
extended partition, and this seems to be a hack that was put in to make
it easy for LILO to install to one.  Personally I'd prefer going back to
the old behavior of just not having a useless device there.

>  BTW, Linux does not use this policy for the others nested partition
>  tables (e.g. Solaris, BSD, Minix, ...). The extended dos partition
>  table is exception. The primary partitions for the others nested PT
>  are exported to the system with its real size :-)

Indeed, also the hidden space in the logical partitions is also not
exposed, otherwise you would have two dev nodes per logical partition.

Re: Extended partition mapping wrong size

[Karel Zak]

On Tue, Apr 06, 2010 at 09:58:39AM -0400, Phillip Susi wrote:
> On 4/6/2010 7:47 AM, Karel Zak wrote:
> > This is probably kernel bug. It's really insane that the extended
> > pseudo partition overflows to the next logical partition.
> 
> Indeed.
> 
> >  Please no. I think the size should not be more than 2 sectors (1024
> >  bytes). The current concept works for years and we have in userspace 
> >  /etc/partitions parsers that use "if (blocks <= 1)" to detect

Sorry, /proc/partitions

> >  extended partitions.
> 
> Could you elaborate a bit on this?  What programs have such tests and
> what would they do differently if it were larger?

I know about libblkid in e2fsprogs and util-linux-ng. It scans
/proc/partitions to get list of "normal" block devices.

We have no clue how many programs/scripts depend on this behaviour.

> >  The other problem are mkfs programs, the space used for alignment
> >  could be 1MiB (or more) -- it's enough many mkfs programs.
> 
> What's wrong with that? 

Irrelevant question ;-) It's there for years and it's well know kernel
feature. 

I understand that from a pedantic point of view the current solution is
not perfect, but I don't see any practical reason why we need to change
anything. There is no issue. Right?

> If you REALLY want to, there's no reason you can't create a tiny fs there. 

You have to care about the partition table (EBR).

The current 1024 bytes is completely useless size, if you enlarge the
size of the partition (for example to 1MiB) you will see reports from
people who lost their extended partitions. (I don't believe that all
mkfs programs are able to detect/skip EBR.)

> Then again, I could swear that once upon a time the kernel simply
> did not bother creating a dev node for the extended partition, and
> this seems to be a hack that was put in to make it easy for LILO to
> install to one.  Personally I'd prefer going back to the old
> behavior of just not having a useless device there.

This is probably better idea than enlarge the size :-)

    Karel

-- 
 Karel Zak  <kzak@redhat.com>
 http://karelzak.blogspot.com

Re: Extended partition mapping wrong size

[Phillip Susi]

On 4/6/2010 11:33 AM, Karel Zak wrote:
> You have to care about the partition table (EBR).
> 
> The current 1024 bytes is completely useless size, if you enlarge the
> size of the partition (for example to 1MiB) you will see reports from
> people who lost their extended partitions. (I don't believe that all
> mkfs programs are able to detect/skip EBR.)

Good point... mke2fs won't damage it since it leaves the first sector
intact and places its boot sector in sector 1 ( or was it 2? ), but FAT
and NTFS won't be so kind.

>> Then again, I could swear that once upon a time the kernel simply
>> did not bother creating a dev node for the extended partition, and
>> this seems to be a hack that was put in to make it easy for LILO to
>> install to one.  Personally I'd prefer going back to the old
>> behavior of just not having a useless device there.
> 
> This is probably better idea than enlarge the size :-)

Aye, I'd prefer the useless device to be removed, but at the very least
it should not use a size of 2 when the second sector is not actually
there because it is the first sector of the logical partition, though we
just patched parted to reserve that second sector anyhow even though you
are using none alignment mode, just to avoid running into this problem.