[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt] 'build' on FS pool now unconditionally formats?



On 02/26/2010 06:23 AM, Daniel P. Berrange wrote:
On Thu, Feb 25, 2010 at 10:01:09PM -0500, Dave Allan wrote:
On 02/25/2010 08:25 AM, Daniel P. Berrange wrote:
On Thu, Feb 25, 2010 at 01:19:54PM +0100, Daniel Veillard wrote:
On Wed, Feb 24, 2010 at 03:51:45PM -0500, Cole Robinson wrote:
Hi guys,

Looking at the new FS pool build options and talking with Dave, I see
that
calling PoolBuild on an FS pool now unconditionally calls mkfs. This is
really
bad when mixed with virt-manager: previously, we assumed the FS build
command
was always non destructive (at most it created a directory), so we
called it
every time, and didn't even allow users to opt out, since there wasn't a
use
case that called for it.

This new formatting behavior really needs to be opt in, otherwise all
virt-manager versions creating an FS pool can destroy data.

Just FYI, for disk pools (and certain LVM configurations) where this
operation
has always been destructive, we default to build=off, and loudly warn
the user
if they choose otherwise. We can do that with this new option as well,
but the
previous behavior really needs to be reinstated IMO (and before the new
release).

I fully accept that this could be a bug in virt-manager's assumptions of
the
build command, but even consider a virsh user: previously build just
created a
directory, now it formats a partition, without any XML change.

I was initially reluctant of changing the behaviour, and asked to use a
flag to keep the original default semantic. I got convinced that noone
could rely on it because the function was basically incomplete. But since
virt-manager ships with an expectation on the previous behaviour, I
revert my position, we need to add a _FORMAT = 4 flag for this call and
only call mkfs if that flag is passed. Fix is trivial we should not
push 0.7.7 without it,

I agree, we should not push 0.7.7 with the code the way it is now, but
what I'm describing below is also trivial, so it wouldn't push back the
release.

I really don't want to  add an extra flag, because it makes filesystem
pool a special case. The 'build' operation is intentionally destructive
by its very definition, and virt-mnager should never be expecting it to
be safe to call on specific pool types.

The problem is that what build means has never been defined, and while
it may have been the intention to implement only destructive operations,
the backends implement a variety of actions.  For some backends build is
is easy to reverse; others not.  The only guidance virsh help gives is
"build a pool" which doesn't indicate any danger at all.  I would define
build as "make the changes to the media necessary to start the pool" and
split those changes into destructive and non-destructive actions with a
flag.  (see below)

The distinction of destructive vs non-destructive does not make sense
at an API level. The build operation is intended to do whatever data
formatting changes are required in order to construct the pool. Whether
this is destructive or not is really an artifact of whether there is any
data already on the storage that you already care about.

This is why I think the build() operation, with flags=0, should refuse
to run if it detects that the pool already has been constructed. This
will protect against data loss with virt-manager's current usage, and
also avoid changing the semantics of the current LVM/disk backends.

Adding a OVERWRITE flag will then allow apps to explicitly ignore the
checks for existing data, allowing possibly destructive operation.


IMHO, we should do two things to address this

  - Fix virt-manager to not call build all the time for any pool
    type - it must only do it when expkicitly requested

  - Make the 'build' operation check to see if the pool is
    already constructed (eg  LVM magic check for logical pools,
    FAT partition check for disk ools&   filesystem magic check
    for the fs pool). Reject the build operation if any of these
    show that the pool exists / is alread ybuilt
  - Add a 'OVERWRITE' flag, to allow apps to forcably reformat,
     regardless of current state

I propose we add a DESTRUCTIVE flag and require it for destructive
operations on all the backends.  The downside, obviously, is that it
changes the behavior of the disk and LVM backends that currently don't
require a flag for destructive operations.  I'm not too worried about
that behavioral change, though, because what's in the tree right now
changes the behavior of the fs backend making a previously
non-destructive operation into a destructive operation without any
change on the users part and without warning.

I don't want us to be changing the semantics of the existing LVM/disk
backends here. In terms of our release schedule, I think we should just
revert the current change to the FS backend completely. Then make the
release. Then re-apply the change, with some followup patches to add
the checks for existing formatted data in the next release.

Ok, I reverted the original patch as we're all in agreement on that.

The utility of the check is dependent on what's on the disk. If the user specifies the wrong partition, which is not hard to do, for example, a choosing a database partition instead of the intended filesystem in the virt-manager gui, the partition will still be overwritten without warning, whereas other overwrite operations are, as Cole says, loudly warned about. The only way to prevent that is to require the flag before any fs format. It's somewhat gross, I freely admit, but I feel fairly strongly that we need to err on the side of caution when dealing with formatting.

I'm ok with leaving the other backends the way they are.

Dave


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]