Re: API definition for LUKS key management

[Maxim Levitsky]

On Mon, 2019-11-11 at 18:34 +0000, Daniel P. Berrangé wrote:
> On Mon, Nov 11, 2019 at 05:58:20PM +0200, Maxim Levitsky wrote:
> > Hi!
> > 
> > I would like to discuss the API for LUKS key management.
> > 
> > First of all very brief overview of LUKS v1 format:
> > 
> > Each sector of the image is encrypted with same master key, which
> > is not stored directly on the disk.
> > 
> > Instead in the LUKS header we have 8 slots. Each slot optionally stores
> > an encrypted version of the master key, encrypted by the user password.
> > Knowing the password, you can retrieve the master key from the keyslot.
> > Slot can be marked as active or inactive, inactive slots are not considered
> > when opening the image.
> > 
> > In addition to that LUKS header has a hash of the master key, so that
> > you can check if the password 'opens' a keyslot by decrypting it
> > with given the password, and then checking if 
> > the hash of the decrypted master key candidate obtained matches the stored 
> > hash.
> > 
> > That basically means that you can have up to 8 different passwords that will
> > open a luks volume and you can change them as you wish without re-encrypting
> > everything.
> > 
> > Now for raw luks volumes you have cryptsetup which allows to manage these
> > keyslots, but we also have so called encrypted qcow2 format which
> > basically has the luks header, together with keyslots embedded, plus each
> > cluster is encrypted with the master key as in raw luks.
> > Cryptsetup doesn't support this, thus I implemented this in qemu block 
> > layer.
> 
> Even for raw luks volumes, the traditional "cryptsetup" tool is
> undesirable. eg consider LUKS on an RBD or ISCSI volume where
> you are using the in-QEMU RBD/ISCSI client. You don't want to
> have to configure the host kernel client just to change the
> keyslot info. You don't want to use the in-QEMU clients for
> qemu-img.

I didn't thought about it. This is a very good point!

> 
> > 
> > Link to bugzilla here: https://bugzilla.redhat.com/show_bug.cgi?id=1662412
> > 
> > 
> > Relevant to the API,
> > first of all qemu has the notion of amend (qemu-img amend), which allows
> > currently to change format specific extensions of qcow2.
> > 
> > Since luks, especially luks inside qcow2 is a format on its own, it fits to 
> > use that interface to change the 'format' options, in this case,
> > the encryption key slots.
> > 
> > 
> > There are the following requirements (they are 100% hardcoded, we might 
> > discuss
> > to drop some of these):
> > 
> > 
> > 1. ability to add a new password to a free keyslot 
> > (best is to let api to pick a free keyslot)
> > Also user should not need to know all the passwords in existing keyslots.
> > 
> > 
> > 2. ability to erase a keyslot, usually by giving the password that should 
> > be erased, and erasing all
> > the keyslots that match the password, or by giving a keyslot index.
> > This will usually be done after adding a new password.
> > 
> > 
> > 3. Allow to do so online, that is while qemu is running, but also support 
> > offline management.
> > Note that online management is even useful for raw luks volumes, since its 
> > not safe
> > to run cryptsetup on them while qemu is using the images.
> > 
> > 
> > I implemented those requirements using the following interface.
> > (I have sent the patches already)
> > 
> > I will try to explain the interface with bunch of examples:
> > 
> > 
> > # adds a new password, defined by qemu secret 'sec0' to first unused slot
> > # give user a error if all keyslots are occupied
> > qemu-img amend --secret ... -o key-secret=sec1 image.luks
> 
> I think you meant "--object secret,...." instead of "--secret ..."
> 
True, sorry about that!

> Also, this example needs to have 2 secrets provided. The first
> secret to unlock the image using the existing password, and the
> second secret is the one being added.
> 
> > # erases all keyslots that can be opened by password that is contained in a 
> > qemu secret 'sec0'
> > # active=off means that the given password/keyslot won't be active after 
> > the operation
> > qemu-img amend --secret ... -o key-secret=sec0,active=off image.luks
> > 
> > 
> > # erase the slot 5 (this is more low level command, less expected to be 
> > used)
> > qemu-img amend --secret ... -o slot=5,active=off image.luks
> > 
> > # add new secret to slot 5 (will fail if the slot is already marked as 
> > active)
> > qemu-img amend --secret ... -o slot=5,key-secret=sec1 image.luks
> 
> This also needs two secrets provideed.
> 
> > 
> > 
> > This is basically it.
> > 
> > The full option syntax is as following:
> > 
> > active=on/off (optional, default to on) - toggles if we enabling a keyslot 
> > or are erasing it.
> > 
> > slot=number (optional, advanced option) - specifies which exactly slot to 
> > erase or which
> > slot to put the new key on
> > 
> > key-secret = id of the secret object - specifies the secret. when slot is 
> > enabled,
> > it will be put into the new slot. when disabling (erasing a keyslot), all 
> > keyslots
> > matching that secret will be erased. 
> > Specifying both key-secret and slot index is treated as error I think
> > 
> > 
> > As as very advanced option, --force is added to qemu-img to allow to do 
> > unsafe operation,
> > which in this case is removing last keyslot which will render the encrypted 
> > image useless.
> > 
> > 
> > In addition to that, QMP interface was added for online version of the 
> > above.
> > It is very similiar, but since we don't have blockdev-amend,
> > I added one and it has the following interface:
> > 
> > 
> > 
> > ##
> > # @x-blockdev-amend:
> > #
> > # Starts a job to amend format specific options of an existing open block 
> > device.
> > # The job is automatically finalized, but a manual job-dismiss is required.
> > #
> > # @job-id:          Identifier for the newly created job.
> > #
> > # @node-name:       Name of the block node to work on
> > #
> > # @options:         Options (same as for image creation)
> > #
> > # @force:           Allow unsafe operations, format specific
> > #                   For luks that allows erase of the last active keyslot
> > #                   (permanent loss of data),
> > #                   and replacement of an active keyslot
> > #                   (possible loss of data if IO error happens)
> > #
> > # Since: 4.2
> > ##
> > { 'command': 'x-blockdev-amend',
> >   'data': { 'job-id': 'str',
> >             'node-name': 'str',
> >             'options': 'BlockdevCreateOptions',
> >             '*force': 'bool' } }
> > 
> > 
> > 
> > It takes the same BlockdevCreateOptions as blockdev-create (this is open to 
> > debate if to leave this as is)
> > 
> > 
> > BlockdevCreateOptionsLUKS (its parent QCryptoBlockCreateOptionsLUKS 
> > technically is extended in this way):
> > 
> > 
> > --- a/qapi/crypto.json
> > +++ b/qapi/crypto.json
> > @@ -190,6 +190,21 @@
> >  #                  Currently defaults to 'sha256'
> >  # @hash-alg: the master key hash algorithm
> >  #            Currently defaults to 'sha256'
> > +#
> > +# @active: Should the new secret be added (true) or erased (false)
> > +#          (amend only, since 4.2)
> > +#
> > +# @slot: The slot in which to put/erase the secret
> > +#        if not given, will select first free slot for secret addtion
> > +#        and erase all keyslots that match the given @key-secret for erase.
> > +#        except the last one
> > +#        (optional, since 4.2)
> > +#
> > +# @unlock-secret: The secret to use to unlock the image
> > +#        If not given, will use the secret that was used
> > +#        when opening the image.
> > +#        (optional, for amend only, since 4.2)
> > +#
> >  # @iter-time: number of milliseconds to spend in
> >  #             PBKDF passphrase processing. Currently defaults
> >  #             to 2000. (since 2.8)
> > @@ -201,7 +216,12 @@
> >              '*cipher-mode': 'QCryptoCipherMode',
> >              '*ivgen-alg': 'QCryptoIVGenAlgorithm',
> >              '*ivgen-hash-alg': 'QCryptoHashAlgorithm',
> > +
> >              '*hash-alg': 'QCryptoHashAlgorithm',
> > +            '*active' : 'bool',
> > +            '*slot': 'int',
> > +            '*unlock-secret': 'str',
> > +
> >              '*iter-time': 'int'}}
> > 
> > 
> > Here note that key-secret is already present in the in api, and I am adding 
> > the 'slot','active' and 'unlock-secret'
> > 
> > 'slot' can be also used for new created image to specify where to place the 
> > the secret.
> > 'active' not allowed to be false for blockdev-create of an image and can be 
> > true/false for 'blockdev-amend'
> > 
> > 'unlock-secret' (might be removed later) covers an corner case that is 
> > specific for online key management.
> > The case is that if the keyslot used to open the image in first place is 
> > removed, it can be used to specify
> > the password to retrieve the master key from one of existing keyslots, 
> > since the driver doesn't officially
> > keep the master key all the time (it can be in theory only loaded in 
> > hardware crypto device)
> > 
> > That is why for adding a new keyslot, the secret that was used to open the 
> > image is tried first, and if it
> > doesn't open a keyslot, the 'unlock-secret' can be used instead. This can 
> > be thought of as the 'current password'
> > that is need to update the password on many web forums.
> > 
> > 
> > One of the concerns that was raised during the review was that amend 
> > interface for luks that I propose is
> > different from the amend inteface used currently for qcow2.
> > 
> > qcow2 amend interface specifies all the format options, thus overwrites the 
> > existing options.
> > Thus it seems natural to make the luks amend interface work the same way, 
> > that it receive an array
> > of 8 slots, and for each slot specify if it is active, and if true what 
> > password to put in it.
> > This does allow to add and erase the keyslots, but it doesn't allow:
> > 
> >    * add a password without knowing all other passwords that exist in 
> > existing keyslots
> >      this can be mitigated by specifying which keyslots to modify for 
> > example by omitting the
> >      keyslots that shouldn't be touched from the array (passing null 
> > placeholder instead)
> >      but then it already doesn't follow the 'specify all the options each 
> > time' principle.
> 
> I think this is highly undesirable, as we must not assume that the
> mgmt app has access to all the passwords currently set.
> 
> The two key use cases for having multiple key slots are
> 
>   - To enable a two-phase change of passwords to ensure new password
>     is safely written out before erasing the old password
>     
>   - To allow for multiple access passwords with different controls
>     or access to when each password is made available.
> 
>     eg each VM may have a separate "backup password" securely
>     stored off host that is only made available for use when
>     doing disaster recovery.
> 
> the second use case is doomed if you need to always provide all
> current passwords when changing any key slots.
Fully agree, and thanks for these examples!


> 
> 
> >    * erase all keyslots matching a password - this is really hard to do 
> > using this approach,
> >      unless we give user some kind of api to try each keyslot with given 
> > password,
> >      which is kind of ugly and might be racy as well.
> > So what do you think?
> 
> The point of using "amend" is that we already have some of the boilerplate
> supporting framework around that, so it saves effort for both QEMU and
> our users. If the semantics of "amend" don't fit nicely though, then the
> benefit of re-using "amend" is cancelled out and we should go back to
> considering a separate "key-manage" command.

I guess we should anyway use amend interface, while updating its definition
a bit to suit the broader requirements of the drivers, e.g luks.

I see the amend interface being like generic 'edit the device options' thing,
which is maybe better that adding device specific commands like 
add-key,remove-key,
etc. No strong opinion on this though.


Best regards,
        Maxim Levitsky