Message-ID: <20190110151256.GN3599@linux-l9pv.suse>
Date:   Thu, 10 Jan 2019 23:12:56 +0800
From:   joeyli <jlee@...e.com>
To:     Andy Lutomirski <luto@...nel.org>
Cc:     Pavel Machek <pavel@....cz>,
        "Lee, Chun-Yi" <joeyli.kernel@...il.com>,
        "Rafael J . Wysocki" <rjw@...ysocki.net>,
        LKML <linux-kernel@...r.kernel.org>, linux-pm@...r.kernel.org,
        keyrings@...r.kernel.org,
        "Rafael J. Wysocki" <rafael.j.wysocki@...el.com>,
        Chen Yu <yu.c.chen@...el.com>,
        Oliver Neukum <oneukum@...e.com>,
        Ryan Chen <yu.chen.surf@...il.com>,
        David Howells <dhowells@...hat.com>,
        Giovanni Gherdovich <ggherdovich@...e.cz>,
        Randy Dunlap <rdunlap@...radead.org>,
        Jann Horn <jannh@...gle.com>
Subject: Re: [PATCH 0/5 v2][RFC] Encryption and authentication for hibernate
 snapshot image

On Wed, Jan 09, 2019 at 10:47:42AM -0800, Andy Lutomirski wrote:
> On Wed, Jan 9, 2019 at 8:40 AM joeyli <jlee@...e.com> wrote:
> >
> > Hi Andy,
> >
> > Thanks for your review!
> >
> > On Tue, Jan 08, 2019 at 01:41:48PM -0800, Andy Lutomirski wrote:
> > > > On Jan 7, 2019, at 9:37 AM, joeyli <jlee@...e.com> wrote:
> > > >
> > > > Hi Pavel,
> > > >
> > > > Thanks for your review!
> > > >
> > > >> On Sun, Jan 06, 2019 at 07:10:27PM +0100, Pavel Machek wrote:
> > > >> Hi!
> > > >>
> > > >>> This patchset is the implementation of encryption and authentication
> > > >>> for hibernate snapshot image. The image will be encrypted by AES and
> > > >>> authenticated by HMAC.
> > > >>
> > > >> Ok, so you encrypt.
> > > >
> > > > Yes, encryption and authentication.
> > > >
> > > >>> The hibernate function can be used to snapshot memory pages to an image,
> > > >>> then kernel restores the image to memory space in a appropriate time.
> > > >>> There have secrets in snapshot image and cracker may modifies it for
> > > >>> hacking system. Encryption and authentication of snapshot image can protect
> > > >>> the system.
> > > >>>
> > > >>> Hibernate function requests the master key through key retention service.
> > > >>> The snapshot master key can be a trusted key or a user defined key. The
> > > >>> name of snapshot master key is fixed to "swsusp-kmk". User should loads
> > > >>> swsusp-kmk to kernel by keyctl tool before the hibernation resume.
> > > >>> e.g. The swsusp-kmk must be loaded before systemd-hibernate-resume
> > > >>
> > > >> But if userspace has a key, encryption is useless against root.
> > > >
> > > > Yes, but this concern is not only for hibernation encryption. This patch
> > > > set does not provide solution against this concern.
> > > >
> > > > The purpose of this patch set is to encrypt and authenticate hibernate
> > > > snapshot image in kernel space. It also requests key through keyring
> > > > mechanism. Which means that we can easy to adapt to new key type from
> > > > keyring in the future.
> > > >
> > > > Currently TPM trusted key or user defined key types are not against
> > > > root. Even using the TPM trusted key, it still can be unsealed by root
> > > > before the PCRs be capped (unless we capped PCRs in kernel).
> > > >
> > > > My solution for keeping the secret by kernel is the EFI secure key type:
> > > >    https://lkml.org/lkml/2018/8/5/31
> > > >
> > > > But the EFI boot variable doesn't design for keeping secret, so Windows
> > > > and OEM/ODM do not use boot variable to keep secret. So this idea can
> > > > not be accepted. We must think other key type against root.
> > > >
> > > >>> The TPM trusted key type is preferred to be the master key. But user
> > > >>> defined key can also be used for testing or when the platform doesn't
> > > >>> have TPM. User must be aware that the security of user key relies on
> > > >>> user space. If the root account be compromised, then the user key will
> > > >>> easy to be grabbed.
> > > >>
> > > >> In the TPM case, does userland have access to the key?
> > > >
> > > > In the TPM case, userland can only touch the sealed key blob. So userland
> > > > doesn't know the real secret. But it has risk that root unseals the key
> > > > before PCRs be capped.
> > > >
> > > >> Please explain your security goals.
> > > >
> > > > My security goals:
> > > >
> > > > - Encrypt and authicate hibernate snapshot image in kernel space. Userspace
> > > >  can only touch an encrypted and signed snapshot image.
> > >
> > > Signed?
> > >
> > > I’m not entirely convinced that the keyring mechanism is what you
> > > want. ISTM that there are two goals here:
> > >
> > > a) Encryption: it should be as hard as can reasonably be arranged to
> > > extract secrets from a hibernation image.
> > >
> > > b) Authentication part 1: it should not be possible for someone in
> > > possession of a turned-off machine to tamper with the hibernation
> > > image such that the image, when booted, will leak its secrets. This
> > > should protect against attackers who don’t know the encryption key.
> > >
> > > c) Authentication part 2: it should be to verify, to the extent
> > > practical, that the image came from the same machine and was really
> > > created using hibernation.  Or maybe by the same user.
> > >
> > > For (a) and (b), using an AE mode where the key is protected in some
> > > reasonable way.  Joey, why are you using HMAC?  Please tell me you’re
> > > at least doing encrypt-then-MAC.  But why not use a real AE mode like
> > > AES-GCM?
> >
> > The reason for using HMAC is the history for development. My first patch
> > set is only for hibernate authentication. Then I added encryption code on
> > top of my authentication patches in last version.
> >
> > I am doing encrypt-then-MAC. My code ecrypts each page by AES then HMAC
> > whole snapshot image. I feed encrypted data pages one by one to
> > crypto_shash_update() API for calculating the hash for whole image.
> 
> ...
> 
> I think you should write down a clear description of the data format.

Hibernation allocates free pages for building snapshot image. Those free
pages are scattered in memory. So kernel marks those page numbers on a
bitmap to locate them. Because this image buffer is discontinuous, so I
use update mode hashing whole image.

> A general problem with crypto is that the fact that it appears to work
> doesn't mean it's secure at all, and it's very hard to follow the
> code.  Especially in Linux using the crypto API -- code using the
> crypto API tends to be mostly boilerplate.
>

hm... Do you mean that the implementation of HMAC in crypto cannot be
trusted? I hope at least that I can trust the update mode for normal
hash like SHA256 or SHA512?
 
> >
> > >
> > >
> > > I reviewed the code a bit.  Here are some thoughts:
> > >
> 
> >
> > > You’re explicitly checking that it’s not zero, and I don’t see why.
> > >
> >
> > That's because I use trampoline page to forward the key from boot
> > kernel to resume kernel for next hibernation cycle. When resuming,
> > the empty key means that the boot kernel didn't forward valid key
> > to resume kernel. Then the key initial failed, hibernation can not
> > be triggered in next cycle.
> >
> 
> This seems like a poor design.  If you need some indication of "there
> is no key", then implement that.  Don't use a special all-zero value
> to mean something special.
>

OK, I will use other way to detect the state of key. 
 
> >
> > > the acceptable usage of trusted keys?
> > >
> >
> > Sorry for I didn't capture the meaning of "acceptable usage". The trusted
> > key already be unsealed by TPM when the key be enrolled by keyctl tool.
> > So my code just extract the unsealed key data (the random number) for
> > encryption.
> 
> If someone creates a trusted key that is used for asymmetric crypto or
> perhaps a trusted key that is intended to be used for, say, an HMAC
> key, you should not also use it to derive hibernation keys.  This is
> what I mean by "acceptable usage".
>

When keyring is producing encrypted key, the trusted key be used to
derive the encrypt key and authenticate key to encrypt and hmac sign
a encrypted key. So trusted key can be used in symmetric crypto.
 
> >
> > > You are using a non-ephemeral key and generating a fresh IV each time.
> > > This is probably okay, but it’s needlessly fragile. Just generate an
> > > entirely fresh key each time, please.  You also seem to be doing
> > > encrypt-and-MAC, which is not generally considered acceptable. And
> >
> > I have thought about using one time key before. But that means I need
> > attach the key with snapshot image. Keyring API doesn't provide interface
> > for other kernel subsystem to feed an encrypted key blob, so I choice
> > to use non-ephemeral key that's enrolled by userland through keyctl. I
> > can change to one time key if keyring subsystem allows to be modified.
> 
> I don't see what this has to do with the keyring API.  If you want a
> one time key in the hibernation code, generate a one-time key in the
> hibernation code and wrap the key accordingly.
>

When using trusted key, the keyring subsystem helps to deal TPM. On the
other hand, I want that hibernate can easy to support other key types. e.g.
user defined key can also be used when debugging or development.

If hibernation direct deals with TPM to seal the hibernation key. Then
yes, we don't need keyring as the abstract layer for hibernation key.
 
> 
> >
> > The encrypt-and-MAC is not acceptable? I am OK to change to other way
> > just I hope that I can encrypt/decrypt page one by one for the reason
> > that I just mentioned.
> >
> > > you’re not authenticating everything — just the data. This seems very
> > > weak.
> > >
> >
> > Is it too weak? Why?
> >
> > There have three parts of an snapshot image:
> >         image header :: page table pages :: data pages
> >
> > The real data are in data pages. I thought that encrypt/hmac data pages
> > can protect sensitive data in pages and also avoid those pages be arbitrary
> > modified. IMHO arbitrary modifing page table pages is easy to cause
> > resume kernel crashed.
> 
> First you need to define *why* you're authenticating anything.  But I
> would guess that under any reasonable security model, a tampered-with
> image should not be able to take over the system.  And, if you don't
> protect the page tables, then it's probably quite easy to tamper with
> an image such that the HMAC is unaffected but the image takes over the
> system.
>

OK, I will protect page tables and also fields in image header.
 
> I think that, if this series goes in at all, it should authenticate
> the *entire* image.  (But you still have to define what "authentic"
> even means.  "It passed an HMAC check" is not a good definition.)
>

I thought that the authentic of image means that the image should not
be tampered. I will try to protect whole image.
 
> >
> > > Can you explain the trampoline?  It looks like you are using it to
> > > tell the resumed kernel that it was tampered with. If so, NAK to that.
> > > Just abort.
> > >
> >
> > The main job of trampoline page is using by boot kernel to forward
> > key to resume kernel for next hibernation cycle. The authorization result
> > is just an add on.
> >
> > In resume process, the key will be loaded in early boot stage. Either in
> > EFI stub (EFI secure key type, be rejected by EFI subsystem) or initrd. On
> > the other hand, PCRs may capped after TPM trusted key be loaded. So resume
> > kernel has no chance to load or unseal key again. The only way to get the
> > unsealed/decrypted key is from boot kernel.
> >
> > The trampoline page is an empty page that it's pre-reserved in snapshot image
> > when hibernation. When resuming, this trampoline page will be written by boot
> > kernel to fill unsealed key. Then resume kernel will check the trampoline
> > page to take the key. If the key is empty(zero), which means that boot
> > kernel didn't forward valid key because some reasons. Either the key is
> > broken (EFI key be erased) or userspace didn't enroll key to boot kernel.
> > That's why I check the zero key in code.
> 
> I don't follow this at all.  After reading the code, I couldn't figure
> out which stage generates the trampoline, which stage reads the
> trampoline, and where there are any keys at all in the trampoline.
>

Very thinks for your review to my code...
 
> There are effectively three kernels:
> 
> a) The kernel that gets hibernated and generates the hibernation
> image.  This kernel obviously has access to all the keying
> information, although some of that information may actually be
> resident on the TPM and not in memory.
> 
> b) The boot kernel that reads the hibernation image.
> 
> c) The resumed kernel.  This kernel has the same code as (a) but
> possibly different PCRs.
> 
> If the encryption is going to protect anything at all, then kernel (b)

Sorry for I forgot mention in previous mail...

Here is the point. I developed authentication before encryption, so the
image is not encrypted in early version. In my original design, I try
to avoid that hibernate key be snapshotted to image. So I erased
hibernate key from kernel (a), which means that the kernel (b) should
forwards hibernate key to kernel (c). Otherwise next hibernate cycle
will be failed because no key.

You are right. When the snpashot image be encrypted, the key doesn't
need to be erased. Then kernel (b) doesn't need to forward key to kernel
(c). 

I will remove trampoline. But, I still have a question when using PCRs
with hibernate as below...

> needs to have a way to get the key, which is presumably done using the
> TPM, since it's not obvious to me that there's any other credible way
> to do it.  (A passphrase could also make sense, but then you'd just
> put the hibernation image on a dm-crypt volume and the problem is
> solved.)  Kernel (c) shouldn't need to do any crypto, since, by the
> time any code in kernel (c) executes at all, you've already decrypted
> everything and you've effectively committed to trusting the
> hibernation image.
> 
> If something goes wrong when you hibernate a second time, this means
> that the first hibernation screwed up and corrupted something, e.g.
> PCRs.  That should be fixed independently of this series IMO.
> 

For kernel lockdown situation, if PCRs are not capped, which means
that a sealed bundle can be unsealed by root. If root be compromised
then the sealed key can also be compromised.

I was thinking if kernel generates a sealed key when booting, then
kernel capped the PCR with a constant value[1] before the userspace
is available. Then the sealed key can not be unsealed by compromised
root. 

In this case, kernel (b) will do the job to produce the sealed key
before PCR be capped. Then the key must be forwarded by kernel (b)
to kernel (c).

[1]
Using constant value for capping but not random number is for other
subsystem or userland can continue to use the PCR.

Thanks a lot!
Joey Lee

Powered by blists - more mailing lists