Re: [PATCH 02/13] qcrypto-luks: implement encryption key management

[Markus Armbruster <armbru@redhat.com>]

Daniel P. Berrangé <berrange@redhat.com> writes:

> On Thu, Jan 30, 2020 at 03:47:00PM +0100, Markus Armbruster wrote:
>> Daniel P. Berrangé <berrange@redhat.com> writes:
>> 
>> > On Thu, Jan 30, 2020 at 01:38:47PM +0100, Kevin Wolf wrote:
>> >> Am 28.01.2020 um 18:32 hat Daniel P. Berrangé geschrieben:
>> >> > On Tue, Jan 28, 2020 at 05:11:16PM +0000, Daniel P. Berrangé wrote:
>> >> > > On Tue, Jan 21, 2020 at 03:13:01PM +0200, Maxim Levitsky wrote:
>> >> > > > On Tue, 2020-01-21 at 08:54 +0100, Markus Armbruster wrote:
>> >> > > > 
>> >> > > > <trimmed>
>> >> > > > 
>> >> > > > > > +##
>> >> > > > > > +# @LUKSKeyslotUpdate:
>> >> > > > > > +#
>> >> > > > > > +# @keyslot:         If specified, will update only keyslot with this index
>> >> > > > > > +#
>> >> > > > > > +# @old-secret:      If specified, will only update keyslots that
>> >> > > > > > +#                   can be opened with password which is contained in
>> >> > > > > > +#                   QCryptoSecret with @old-secret ID
>> >> > > > > > +#
>> >> > > > > > +#                   If neither @keyslot nor @old-secret is specified,
>> >> > > > > > +#                   first empty keyslot is selected for the update
>> >> > > > > > +#
>> >> > > > > > +# @new-secret:      The ID of a QCryptoSecret object providing a new decryption
>> >> > > > > > +#                   key to place in all matching keyslots.
>> >> > > > > > +#                   null/empty string erases all matching keyslots
>> >> > > > > 
>> >> > > > > I hate making the empty string do something completely different than a
>> >> > > > > non-empty string.
>> >> > > > > 
>> >> > > > > What about making @new-secret optional, and have absent @new-secret
>> >> > > > > erase?
>> >> > > > 
>> >> > > > I don't remember already why I and Keven Wolf decided to do this this way, but I think that you are right here.
>> >> > > > I don't mind personally to do this this way.
>> >> > > > empty string though is my addition, since its not possible to pass null on command line.
>> >> > > 
>> >> > > IIUC this a result of using  "StrOrNull" for this one field...
>> >> > > 
>> >> > > 
>> >> > > > > > +# Since: 5.0
>> >> > > > > > +##
>> >> > > > > > +{ 'struct': 'LUKSKeyslotUpdate',
>> >> > > > > > +  'data': {
>> >> > > > > > +           '*keyslot': 'int',
>> >> > > > > > +           '*old-secret': 'str',
>> >> > > > > > +           'new-secret' : 'StrOrNull',
>> >> > > > > > +           '*iter-time' : 'int' } }
>> >> > > 
>> >> > > It looks wierd here to be special casing "new-secret" to "StrOrNull"
>> >> > > instead of just marking it as an optional string field
>> >> > > 
>> >> > >    "*new-secret": "str"
>> >> > > 
>> >> > > which would be possible to use from the command line, as you simply
>> >> > > omit the field.
>> >> > > 
>> >> > > I guess the main danger here is that we're using this as a trigger
>> >> > > to erase keyslots. So simply omitting "new-secret" can result
>> >> > > in damage to the volume by accident which is not an attractive
>> >> > > mode.
>> >> 
>> >> Right. It's been a while since I discussed this with Maxim, but I think
>> >> this was the motivation for me to suggest an explicit null value.
>> 
>> A bit of redundancy to guard against catastrophic accidents makes sense.
>> We can discuss its shape.
>> 
>> >> As long as we don't support passing null from the command line, I see
>> >> the problem with it, though. Empty string (which I think we didn't
>> >> discuss before) looks like a reasonable enough workaround to me, but if
>> >> you think this is too much magic, then maybe not.
>> >> 
>> >> > Thinking about this again, I really believe we ought to be moire
>> >> > explicit about disabling the keyslot by having the "active" field.
>> >> > eg
>> >> > 
>> >> > { 'struct': 'LUKSKeyslotUpdate',
>> >> >   'data': {
>> >> >           'active': 'bool',
>> >> >           '*keyslot': 'int',
>> >> >           '*old-secret': 'str',
>> >> >           '*new-secret' : 'str',
>> >> >           '*iter-time' : 'int' } }
>> >> > 
>> >> > "new-secret" is thus only needed when "active" == true.
>> 
>> I figure @iter-time, too.
>> 
>> >> Hm. At the very least, I would make 'active' optional and default to
>> >> true, so that for adding or updating you must only specify 'new-secret'
>> >> and for deleting only 'active'.
>> >
>> > Is that asymmetry really worth while ? It merely saves a few
>> > characters of typing by omitting "active: true", so I'm not
>> > really convinced.
>> >
>> >> > This avoids the problem with being unable to specify a null for
>> >> > StrOrNull on the command line too.
>> >> 
>> >> If we ever get a way to pass null on the command line, how would we
>> >> think about a struct like this? Will it still feel right, or will it
>> >> feel like we feel about simple unions today (they exist, we would like
>> >> to get rid of them, but we can't because compatibility)?
>> >
>> > Personally I really don't like the idea of using "new-secret:null"
>> > as a way to request deletion of a keyslot. That's too magical
>> > for an action that is so dangerous to data IMhO.
>> 
>> I tend to agree.  Expressing "zap the secret" as '"new-secret": null' is
>> clever and kind of cute, but "clever" and "cute" have no place next to
>> "irrevocably destroy data".
>> 
>> > I think of these operations as activating & deactivating keyslots,
>> > hence my suggestion to use an explicit "active: true|false" to
>> > associate the core action being performed, instead of inferring
>> > the action indirectly from the secret.
>> >
>> > I think this could lend itself better to future extensions too.
>> > eg currently we're just activating or deactivating a keyslot.
>> > it is conceivable in future (LUKS2) we might want to modify an
>> > existing keyslot in some way. In that scenario, "active" can
>> > be updated to be allowed to be optional such that:
>> >
>> >  - active: true ->  activate a currently inactive keyslot
>> >  - active: false -> deactivate a currently active keyslot
>> >  - active omitted -> modify a currently active keyslot
>> 
>> A boolean provides two actions.  By making it optional, we can squeeze
>> out a third, at the price of making the interface unintuitive: how would
>> you know what "@active absent" means without looking it up?
>> 
>> Why not have an @action of enum type instead?  Values "add" and "delete"
>> now (or "activate" and "deactivate", whatever makes the most sense when
>> writing the docs), leaving us room to later add whatever comes up.
>
> I probably worded my suggestion badly - "active" should not be
> thought of as expressing an operation type; it should be considered
> a direct reflection of the "active" metadat field in a LUKS keyslot
> on disk.
>
> So I should have described it as:
>
>  - active: true|false ->  set the keyslot active state to this value
>  - active omitted -> don't change the keyslot active state
>
> The three possible states of the "active" field then happen to
> provide the way to express the desired operations.
>
>> 
>> This also lets us turn LUKSKeyslotUpdate into a union.
>> 
>> Brief detour before I sketch that: update safety.
>> 
>> Unless writing a keyslot is atomic, i.e. always either succeeds
>> completely, or fails without changing anything, updating a slot in place
>> is dangerous: you may destroy the old key without getting your new one
>> in place.
>> 
>> To safely replace an existing secret, you first write the new secret to
>> a free slot, and only when that succeeded, you delete the old one.
>> 
>> This leads to the following safe operations:
>> 
>> * "Activate": write a secret to a free keyslot (chosen by the system)
>> 
>> * "Deactivate": delete an existing secret from all keyslots containing
>>   it (commonly just one)
>> 
>> Dangerous and unwanted:
>> 
>> * Replace existing secret in place
>> 
>> Low-level operations we may or may not want to support:
>> 
>> * Write a secret to specific keyslot (dangerous unless it's free)
>> 
>> * Zap a specific keyslot (hope it contains the secret you think it does)
>> 
>> Now let me sketch LUKSKeyslotUpdate as union.  First without support for
>> the low-level operations:
>> 
>>     { state: 'LUKSKeyslotUpdateAction',
>>       'data': [ 'add', 'delete' ] }
>>     { 'struct': 'LUKSKeyslotAdd',
>>       'data': { 'secret': 'str',
>>                 '*iter-time': 'int' } }
>>     { 'struct': 'LUKSKeyslotDelete',
>>       'data': { 'secret': 'str' }
>>     { 'union: 'LUKSKeyslotUpdate',
>>       'base': { 'action': 'LUKSKeyslotUpdateAction' }
>>       'discriminator': 'action',
>>       'data': { 'add': 'LUKSKeyslotAdd' },
>>               { 'delete': 'LUKSKeyslotDelete' } }
>> 
>> Since @secret occurs in all variants, we could also put it in @base
>> instead.  Matter of taste.  I think this way is clearer.  Lets us easily
>> add a variant that doesn't want @secret later on (although moving it
>> from @base to variants then would be possible).
>
>
> This kind of approach is what I originally believed we
> should do, but it is contrary to the design expectations
> of the "amend" operation. That is not supposed to be
> expressing operations, rather expressing the desired
> state of the resulting disk.

I got that now, so let's talk state.

A keyslot can be either inactive or active.

Let's start low-level, i.e. we specify the slot by slot#:

    state       new state   action
    inactive    inactive    nop
    inactive    active      put secret, iter-time, mark active
    active      inactive    mark inactive (effectively deletes secret)
    active      active      in general, error (unsafe update in place)
                            we can make it a nop when secret, iter-time
                                remain unchanged
                            we can allow unsafe update with force: true

As struct:

    { 'struct': 'LUKSKeyslotUpdate',
      'data': { 'active': 'bool',       # could do enum instead
                'keyslot': 'int',
                '*secret': 'str',       # present if @active is true
                '*iter-time': 'int' } } # absent if @active is false

As union:

    { 'enum': 'LUKSKeyslotState',
      'data': [ 'active', 'inactive' ] }
    { 'struct': 'LUKSKeyslotActive',
      'data': { 'secret': 'str',
                '*iter-time': 'int } }
    { 'union': 'LUKSKeyslotAmend',
      'base': { 'state': 'LUKSKeyslotState' }   # must do enum
      'discriminator': 'state',
      'data': { 'active': 'LUKSKeyslotActive' } }

When we don't specify the slot#, then "new state active" selects an
inactive slot (chosen by the system, and "new state inactive selects
slots by secret (commonly just one slot).

New state active:

    state       new state   action
    inactive    active      put secret, iter-time, mark active
    active      active      N/A (system choses inactive slot)

New state inactive, for each slot holding the specified secret:

    state       new state   action
    inactive    inactive    N/A (inactive slot holds no secret)
    active      inactive    mark inactive (effectively deletes secret)

As struct:

    { 'struct': 'LUKSKeyslotUpdate',
      'data': { 'active': 'bool',       # could do enum instead
                '*keyslot': 'int',
                '*secret': 'str',       # present if @active is true
                '*iter-time': 'int' } } # absent if @active is false

As union:

    { 'enum': 'LUKSKeyslotState',
      'data': [ 'active', 'inactive' ] }
    { 'struct': 'LUKSKeyslotActive',
      'data': { 'secret': 'str',
                '*iter-time': 'int } }
    { 'union': 'LUKSKeyslotAmend',
      'base': { '*keyslot': 'int',
                'state': 'LUKSKeyslotState' }
      'discriminator': 'state',
      'data': { 'active': 'LUKSKeyslotActive' } }

Union looks more complicated because our union notation sucks[*].  I
like it anyway, because you don't have to explain when which optional
members aren't actually optional.

Regardless of struct vs. union, this supports an active -> active
transition only with an explicit keyslot.  Feels fine to me.  If we want
to support it without keyslot as well, we need a way to specify both old
and new secret.  Do we?


[*] I hope to fix that one day.  It's not even hard.