Re: reboot problem after altering the disk label

[Andrew Clausen]

Matt Wilson wrote:
> 
> fdisk does nothing different than parted when it comes to BLKRRPART.
> You can use parted to modify partition tables on a disk that has a use
> count 1 (one rw fd pointing to the disk, no mounted partitions, etc.)
> In this case the BLKRRPART will succeed and the ll_blk system will
> have the new geometries loaded.
> 
> If the disk is being used, you still have to reboot.

Yep, right.

I should mention, 2.4 has a new blkpg interface, which
allows you to update the kernel's idea of the partition
table incrementally.  So, if only / is mounted, you can
still tell it about the resized swap, or whatever.
(the interface allows you to add and remove partitions)

I've been meaning to implement this for ages...

One dilemma: I think this is more an "OS" issue than
a partition table issue.  So, I think the code for blkpg
should be in the same place as the RRPART ioctl.

It would basically look like:

static void
set_kernel_part_entry (PedDevice* dev, PedPartition* part)
{
        ioctl (dev->fd, BLKPG_REMOVE, part->num);
        ioctl (dev->fd, BLKPG_ADD, {part->num, ...});
}

static void
update_part_table (PedDevice* dev)
{
        disk = ped_disk_open(dev);
        for (part ...) {
                if (!ped_partition_is_active(part)) continue;
                set_kernel_part_entry (part);
        }
        /* remove deleted partitions?  why bother?  Overlapping
         * might cause issues?  Anyway, trivial enough to do ;-)
         * traverse the kernel partition table, and try
         *      ped_disk_get_partition(disk, N)...
         */
        ped_disk_close();
}

See the recent thread on linux-fsdevel, BTW.  An interesting
result of this would is it allows us to remove partition table
support from the kernel.  Just have Parted parse it from an
initrd ;-)

I'm still not sure if that application is a good idea...
(initrd's don't excite me)

To implement the above code in 1.5.x, I think the
kernel_set_part_entry() should be kernel-specific, and
the update_part_table() should be public, and abstract.

(i.e. ped_device_update_kernel_table() ?)

Andrew Clausen