fdisk(8) no longer capable of altering geometry

[Chris Pressey]

[this is a follow-up to / correction of my post to geom@ a few days ago,
to which there was no reply]

Hello,

It appears that fdisk(8) is no longer capable of altering the geometry
of a disk.  (By which I mean, the kernel's idea of the BIOS'es idea of
the geometry, of course.)  I'd find it reassuring to know whether or not
anyone else is seeing the same behaviour, before I go the official route
and file a PR.

Initially I thought that this failure case was only for uninitialized
disks, but I have tried further tests and I can't get fdisk(8) to change
the geometry in any of the cases.  These cases are:

a. uninitialized, totally blank disk (dd if=/dev/zero of=/dev/ad1 ...)
b. initialized disk with FreeBSD (or any other OS) installed on it
c. same as b, but with its root partition mounted on /mnt
d. the disk containing the currently booted FreeBSD system
   (mounted on /, of course)

In each of these cases, I tried a sequence like the following:

  fdisk -BI ad1
  fdisk -u ad1
  <when asked if I want to change "our idea of what the BIOS thinks",
   say yes, and plug in different but compatible values for cylinders,
   heads, and sectors/track>
  <when asked if I want to write the new partition table, say yes>
  fdisk ad1
  <observe that the geometry hasn't changed>

The behaviour I see is, in summary:

a & b: fdisk issues the warning "fdisk: Geom not found" which presumably
       refers to the fact that there is no GEOM MBR provider for that
       disk. It then falls back to the legacy behaviour of raw-writing
       the partition table into sector zero of the disk.  This does not,
       however, trigger an update of the kernel's idea of the geometry.

c & d: no "Geom not found" warning, but no change in geometry either.

I don't see this behaviour on DragonFly; cases a and b work as you would
logically expect (as they worked in 4.x, AFAIR, but I have not yet
tested this) where the geometry does get changed, and subsequent runs of
fdisk report the changed geometry.

In cases c & d, the behaviour is the same as FreeBSD - nothing changes. 
This is not too surprising, since the disk _is_ in use - but an error
message would probably make more sense.

This bug is one of the few remaining things standing in the way of
porting the BSD Installer to FreeBSD.  Without some way of altering the
the geometry, it can't install onto a system whose BIOS misreports the
disk geometry.

My analysis of the problem can be found in my previous post to geom@,
but to sum it up: I think fdisk needs to inform GEOM somehow that the
geometry should be changed.  It might need to trigger the creation of a
GEOM MBR provider for the disk before it does so; but I'm not certain of
any of this, since my knowledge of GEOM is slim at best.

Hopefully someone more familiar with GEOM and such under -CURRENT is
listening and can shed more light on this problem and/or provide a
workaround and/or explain how I'm wrong and show me the right way to do
what I'm trying to do (change geometry) in -CURRENT.

Thanks for your time,
-Chris