Use "real" disks

[clayg]

Mostly the changes should be to the data recipe [1]

I have some code in another context that does this:

@fab.roles('node')
@fab.parallel
def _clean_devices():
    for i in range(DEVICE_COUNT):
        fab.run('dmsetup remove dev%s || true' % i)
        fab.run('losetup -d /dev/loop%s || true' % i)
        fab.run('rm /var/dev%s.block || true' % i)
    fab.run('udevadm settle')


@fab.roles('node')
@fab.parallel
def _create_devices():
    for i in range(DEVICE_COUNT):
        fab.run('dd if=/dev/zero of=/var/dev%s.block bs=1M count=3000' % i)
    for i in range(DEVICE_COUNT):
        fab.run('losetup /dev/loop{0} /var/dev{0}.block'.format(i))
    for i in range(DEVICE_COUNT):
        sectors = fab.run('blockdev --getsize /dev/loop%s' % i)
        fab.run('dmsetup create dev{0} --table '
                '"0 {1} linear /dev/loop{0} 0"'.format(i, sectors))

@fab.task
@fab.runs_once
def setup_devices():
    """
    Add Device Mapper block devices to Nodes
    """
    fab.execute(_clean_devices)
    fab.execute(_create_devices)

... but I think that strategy could be improved further because vsaio's can potentially want many more swift devices than /dev/loop* entries ...

I might just create a single sparse file and loopback that onto /dev/loop0 (like the recipe already does), but then break that up into <sectors> / <DEVICE_COUNT> "chunks" and then use dmsetup to create dev<N> device tables like 0 <chunks> linear /dev/loop0 <N * chunks> [2]

1. https://github.com/swiftstack/vagrant-swift-all-in-one/blob/master/cookbooks/swift/recipes/data.rb
2. probably need to double check that, I think the syntax is <start-of-table-entry> <size-of-table-entry> linear <source-device> <source-device-offset> - since we only create one table entry per device, and always use the same loop device - I think it's always 0 <size-of-each-dev-in-sectors> linear /dev/loop0 <start-sector-offset-of-loop-device> - https://www.kernel.org/doc/Documentation/device-mapper/linear.txt

[tipabu]

Closed by #77