Hi Raghupathy and Sumit,
The Gerrit server gives a "Error 500 (Server Error): Internal server error" today, so my comments here are without having looked at the code. Having said that, I was involved in early discussions regarding this a couple of months ago, so I have some knowledge about all this.
On Fri, 24 Jan 2020 at 19:38, Raghupathy Krishnamurthy < raghu.ncstate@icloud.com> wrote:
Thanks Sumit. We are beginning to go off topic and i don't think you and i agree entirely. I'll lay out my concerns, to see if others on the list share my opinion/concerns, so i hope you can let them respond. If nobody else shares my concerns or responds, the patch will be merged soon anyway, so no point dragging this out :) I apologize for the long email. I don't know how to keep this topic short and detailed. :)
Here are my concerns:
- The patch set claims to implement R060_TBBR_FUNCTION from the TBBR
spec. I don't agree with this. The patch is using authenticated encryption(which does provide confidentiality) but is different from implementing firmware encryption as described in the TBBR(and PSA) in my view. If we are implementing R060_TBBR_FUNCTION, in my view, this should be implemented such that the firmware is first encrypted(symmetric), and the encrypted firmware should be asymmetrically signed(as per PSA). Also, the order for verification should be asymmetric signature verification and then decryption(symmetric).
I think we have to be clear about what we are trying to protect here. Traditional signatures on raw binaries are there so we can be sure that we're loading unmodified firmware coming from the one owning a private key corresponding to the public key hardcoded into the device (or via a hash of the public key). I think we all can agree that a plain signature and traditional CoT fulfills that requirement. Next step is that we want to make it harder to do reverse engineering, how? The idea here is to use encryption. So, then naturally there are questions like, in what order should things happen? Sign-then-encrypt or encrypt-then-sign? Where to store the private key? Do we need integrity protection as well etc? In previous mails in this thread where it's referred to an article that talks about this, they argue that one basically need to do sign-encrypt-sign and in other places on the net people argue that one should do sign-then-encrypt. Here we have to pay attention to the different use cases. We are not trying to protect messages in transit where MITM can strip a signature and replace it with his/hers own signature. We have to remember that we have a fixed set of public keys used to verify the signature and we don't randomly trust arbitrary public keys. I.e., as soon as someone tries to tamper with the signature, it'll be detected due to CoT (originating from immutable secrets/public keys). I think I'd argue that both encrypt-then-sign and sign-then-encrypt would work in our use case. Because of that, I think doing the verification first and then if that is OK do the decryption are quite natural steps to take. If the signature verification fails, well, then someone has tampered with the data and we don't even have to bother trying to decrypt the firmware. I do see that one can turn that argument around, especially when using the authenticated encryption. If the tag is not valid, well, then no decryption and no point in trying to verify the signature (you cannot even do that). My point is that _if_ the use case allows it, then it's better to verify the signature first so you know that after that point you are working with untampered data.
If we ask ourselves the following questions; can an attacker make the devices run tampered firmware in our use case? Can the attacker easily reverse engineer the binary in our use case? I think the answer is no in both cases (actually for both encrypt-then-sign and sign-then-encrypt).
In this patch, plain text firmware is asymmetrically signed, plain text firmware is encrypted and authenticated with a symmetric key using authenticated encryption. The order for verification in this patch is authenticate and decrypt using the symmetric key and THEN asymmetric signature verification on the decryted(and now plain text) firmware. The ordering of verification and decryption are subtly different from what the spec(TBBR and PSA) expects. Security does not *seem* to be broken as far as i can tell, but the patch is not strictly an implementation of the spec, as it claims. This can throw off people reading the spec and trying to match code to spec. This is not good when someone wants to audit and gain confidence in the security of the implementation. At a minimum, we need someone writing the spec to address this.
- The TBBR spec does not talk about authenticated encryption of firmware
and only talks about encryption of firmware. It also does not specify if the encrypted firmware needs to be signed or the plain text need to be signed. This is critical detail. It also does not specify the order in which encryption and signing, decryption and verification must be done. These are important details and need to be addressed explicitly(perhaps in both TBBR and PSA). Use of authenticated encryption needs to be explicitly discussed as well.
Indeed, it'd be good if specs where a bit more precise about that. One reason why we're having this discussion is simply because it isn't clear enough (also, probably too many "may" in the spec?).
- TBBR specifies that the TOC(table of contents) *may* be authenticated.
The TOC contains bits that dictate security policy, in this case, if the firmware is encrypted or not. I don't think it is good to consume security policy from unauthenticated data. In my view, this information must be in the signed image manifest(certificate in ATF) since an attacker can at-least cause simple DoS attacks by flipping bits in the FIP header, undetected. If the bit is flipped, a firmware image going out to a million devices may not boot, since an attacker decided to flip a bit, and signature verification fails, since the boot loader decided not decrypt the firmware based in an unsigned bit of information. PSA appears to have a better approach and makes no provision such data to be outside the image manifest, which is always signed.
I think I'm with you here Raghupathy. Either devices will be configured to run with signed binaries or with both encrypted and signed binaries. Only reason to allow running an unauthenticated header would be in debug builds used in pre-production.
- This last one is a matter of opinion and can live with the current
design. The FIP layer is now aware of the crypto module. FIP has also become coupled with authenticated encryption(If we do this, why not have the FIP layer call auth_mod_verify_img() and why stop at cyrpto_mod_auth_decrypt()) . FIP is security aware. I'd like to keep things the way they are today, where IO and security are separate modules and glued together by another layer(load_auth_image() for example). This is cleaner in my view since security policy(should i decrypt? should i verify signatures? should i apply any security at all?) is separate from IO.
Thanks Raghu
Lastly, am I happy with Sumit's patches? Don't know since I haven't reviewed them. Based on the discussions I believe that they fulfill what we are trying to achieve ("running" encrypted and signed images). But, it seems like there is room for some improvements, like encrypting ToC, agree on sign/verification encrypt/decrypt order and separately get better feedback on the TBBR spec from Arm.
// Regards Joakim
On January 24, 2020 at 3:06 AM, Sumit Garg sumit.garg@linaro.org wrote:
On Fri, 24 Jan 2020 at 04:02, Raghupathy Krishnamurthy raghu.ncstate@icloud.com wrote:
I also just realized that both the TBBR and ARM PSA only talk about encryption of the image, and not authenticated encryption. The guarantees provided by both are completely different. Your review( https://review.trustedfirmware.org/c/TF-A/trusted-firmware-a/+/2495/) talks about the requirement R060_TBBR_FUNCTION being implemented, which is technically not true(and potentially misleading). We must make a note of this difference and use the appropriate terminology, without mixing the two, in the documentation, commit messages, source code comments and error prints. The tool is also called 'encrypt_fw ' but should maybe be named appropriately to indicate it is doing authenticated encryption.
I wouldn't call it misleading. Since firmware encryption feature essentially provides confidentiality protection and authenticated encryption is the type of crypto algorithm which we have used to implement it.
BTW, ARM PSA(file:///home/raghu/repos/fvp/DEN0072-PSA_TBFU_1-0-REL.pdf) expects that the image manifest(X509 certificate) contain the hash of the ENCRYPTED image(Table 2 and as described in my answer to your question "How would this ensure integrity of ciphertext"). The TBBR spec completely misses this fact, and is a crucial detail if we only implement encryption(as opposed to authenticated encryption).Build_macros.mk, in your change, passes the un-encrypted image to cert-tool. You can get away with it in your implementation, since you are using authenticated encryption, not if you were only implementing firmware encryption.
I have already highlighted the issue with signing the ciphertext in my previous reply which deviates from security properties provided by signature verification of plain firmware. So I think we need to revisit ARM PSA TBFU spec.
-Sumit
Is it possible for somebody from ARM to have the TBBR spec updated to reflect this? Also perhaps talk to the spec writers about incorporating authenticated encryption into TBBR and PSA? This patch set is somewhat trailblazing in this regard.
-Raghu
On January 23, 2020 at 12:08 PM, Raghupathy Krishnamurthy via TF-A < tf-a@lists.trustedfirmware.org> wrote:
Hi Sumit,
Thanks for your response.
So firstly I would suggest you to revisit TBBR spec [1],
[RK] I'm very familiar with the TBBR spec and the requirements. Note that not all SoC's adhere perfectly to the TBBR spec, since it does not apply to devices in all market segments. However, these devices do use arm trusted firmware and TBBR CoT in a slightly modified form, which is still perfectly valid. Also, the TBBR spec can be changed if required :)
Why would one use authenticated decryption only to establish TBBR
Chain of Trust providing device the capability to self sign its firmware?
[RK] Fair point. However, you may have devices that don't have the processing power or hardware budget or cost factors(paying for HSM's to store private asymmetric keys), to implement asymmetric verification, in which case using authenticated decryption to verify firmware authenticity and integrity is perfectly valid. The attacks on devices that use symmetric keys to verify firmware authenticity and integrity are usually related to exploiting firmware flaws that leak the key or insiders leaking keys, but that is a different problem and requires different solutions. Fundamentally, there is nothing wrong with using symmetric keys for this purpose, so long as the key is well protected. Also note, security requirements and guarantees are different for different systems. The risk is taken by the system designer and should not be imposed by framework code. I don't advocate doing this but it is an option that your implementation does not provide(and perhaps rightly so).
How would this ensure integrity of ciphertext?
[RK] You sign the ciphertext. In your design, you pass bl31_enc.bin to cert_tool to sign. You don't decrypt the encrypted cipher text until you have verified the asymmetric signature(which provides integrity). As far as signature verification is concerned, whether you sign the plain text or ciphertext is immaterial, since you are simply verifying that the absolute bits you loaded have not been modified(assuming you use a secure signature scheme).
Have a look at some defective sign and encrypt techniques here [2]
[RK] Again, very familiar with [2]. In the S/MIME case, you have multiple parties. With secure boot, you have one party, effectively verifying its own messages across time. There is only one key used to verify signatures. 1.1 and 1.2 does not apply. Also you are encrypting and signing with completely different keys and algorithms. Section 1.2 applies when you use RSA/El-gamal encryption. Here you use symmetric encryption and asymmetric signing.
Why would one not use TBBR CoT here?
[RK] see above. Not all systems are designed equal.
and why would one like to hardcode in a device during
provisioning to boot only either an encrypted or a plain firmware
image?
[RK] Why would you not? You typically want to have the same security policy for a class of devices and not be modifiable by an attacker. It isn't common for the same class of devices to use encrypted firmware some times, and un-encrypted firmware other times. If it is common, there is no problem with setting the bit in the FIP header, as long as verified boot is mandatory. The only concern(as my original email said) is the coupling of the FIP layer and the crypto module, in the implementation. I still don't like that fact that the bit saying the file is encrypted is not signed and this may require talking to the TBBR spec writers. Page 22 of the TBBR spec calls out ToC as "Trusted Table of Contents". The FIP header cannot be "trusted", if it is not in ROM or its integrity is verified by some means! R010_TBBR_TOC should perhaps be mandatory then. Also see R080_TBBR_TOC that says the TOC MUST be ROM'ed or tied by hardware in readable registers. This requirement seems contradictory to R010_TBBR_TOC, given that the FIP header(TOC) is copied from mutable NVM by ROM or some boot stage and then ROM'd or loaded into registers. I may be misunderstanding R080_TBBR_TOC, but i'd interpret it as the TOC(FIP header in ATF implementation of TBBR) as being in ROM or integrity verified.
How would one handle a case where BL31 is in plain format and BL32 is in encrypted format?
[RK]TBBR CoT is equipped to do this. The table is defined on a per image basis.
If you are really paranoid about authentication of FIP header...
[RK] I don't mean to pontificate but there are real world customers buying real hardware, running ATF code, who care about such details and ask about such things routinely. It is not just me being paranoid and is definitely not a minor matter to think of such details. We should discuss more and consider the implications of R080_TBBR_TOC and R010_TBBR_TOC, perhaps on a different thread, without blocking your code review. Can somebody from ARM clarify these requirements with the spec writers?
Thanks
-Raghu
On January 23, 2020 at 12:38 AM, Sumit Garg sumit.garg@linaro.org wrote:
Hi Raghu,
I guess you have completely misunderstood this feature. This is an
optional feature which allows to load encrypted FIP payloads using
authenticated decryption which MUST be used along with signature
verification (or TBBR CoT).
So firstly I would suggest you to revisit TBBR spec [1], especially
requirements: R040_TBBR_TOC, R060_TBBR_FUNCTION etc.
On Thu, 23 Jan 2020 at 00:14, Raghupathy Krishnamurthy
raghu.ncstate@icloud.com wrote:
Hello,
The patch stack looks good. The only comment i have is that the FIP layer has now become security aware and supports authenticated decryption(only). This is a deviation from the secure/signed/verified boot design, where we use the TBBR COT to dictate the security operations on the file. This is nice, because file IO is decoupled from the security policy. This may be a big deviation(i apologize if this was considered and shot down for some other reason), but it may be worthwhile to consider making authenticated decryption a part of the authentication framework as opposed to coupling it with the FIP layer.
It looks like you have mixed both TBBR CoT and this authenticated
decryption feature. They both are completely different and rather
complement each other where TBBR CoT establishes
secure/signed/verified boot and this authenticated decryption feature
provides confidentiality protection for FIP payloads.
At a high level, this would mean adding a new authentication method(perhaps AUTH_METHOD_AUTHENTICATED_DECRYPTION), and having the platform specify that the image is using authenticated encryption in the TBBR COT.
Why would one use authenticated decryption only to establish TBBR
Chain of Trust providing device the capability to self sign its
firmwares? We must use signature verification for TBBR CoT (see
section: 2.1 Authentication of Code Images by Certificate in TBBR spec
[1]).
The authentication framework is already well designed and well equipped to handle these types of extensions.
- This would make the change simpler, since you would not require changes
to the FIP tool and the FIP layer.
- This would also allow for future cases where a platform may want to
only encrypt the file and use public key authentication on the encrypted file(for ex. the soc does not have a crypto accelerator for aes-gcm but only for AES and public key verification, for whatever reason).
How would this ensure integrity of ciphertext? This approach may be
vulnerable to Chosen Ciphertext Attacks (CCAs). Authentication tag as
part of AES-GCM provides integrity protection for ciphertext.
- This would let you choose the order in which you want to do the
authenticated decryption(or just decryption) and signature verification, if you use both, one or the other.
Have a look at some defective sign and encrypt techniques here [2].
The order can't be any arbitrary one, we need to be careful about
this.
One other thing i'm not entirely comfortable with is that the flag indicating if there are encrypted files or not in the FIP, is in the *unsigned* portion of the FIP header. An attacker could simply flip bits that dictate security policy in the header and avoid detection(in this case, the indication that the file needs authenticated decryption). If a platform only uses authenticated encryption, but not verified boot, an attacker could flip the bit in the FIP header and have any image loaded on the platform.
Why would one not use TBBR CoT here?
If authenticated encryption cannot be used without verified boot(which requires build time flags), having a flag to indicate that there are encrypted files in the FIP header is moot, since this can come at build time through the TBBR COT. In any case, it seems like the security policy that firmware images need to be decrypted or authenticated with authenticated decryption, seems like a firmware build time or manufacturing time decision(perhaps a bit set in the e-fuses).
Again you are confusing TBBR CoT with authenticated decryption
feature. And why would one like to hardcode in a device during
provisioning to boot only either an encrypted or a plain firmware
image?
There seems to be no benefit to having a flag in the FIP header.
How would one handle a case where BL31 is in plain format and BL32 is
in encrypted format?
Otherwise, I cant think of any attacks due to this and it may be completely okay, but generally, consuming data that dictates security policy/operations before verifying its integrity seems like a recipe for disaster.
If you are really paranoid about authentication of FIP header then you
should look at implementing optional requirement: R010_TBBR_TOC as per
TBBR spec [1].
[1] https://developer.arm.com/docs/den0006/latest/trusted-board-boot-requirement...
[2] http://world.std.com/~dtd/sign_encrypt/sign_encrypt7.html
-Sumit
-Raghu
On January 22, 2020 at 3:51 AM, Sumit Garg via TF-A < tf-a@lists.trustedfirmware.org> wrote:
Hi Sandrine,
On Wed, 22 Jan 2020 at 15:43, Sandrine Bailleux
Sandrine.Bailleux@arm.com wrote:
Hello Sumit,
Thank you for reworking the patches and addressing all of my review
comments. I am happy with the latest version of these and consider them
ready to go. I plan to leave them in Gerrit for another week to give
extra time for other potential reviewers to have a look and comment.
Thanks for your review.
To everyone on the list: Please raise any concerns you may have about
these patches in the coming week. If I don't hear anything by 29th
January 2020, I will merge these patches.
@Sumit: One of the next actions for this patch stack would be to have
some level of testing in the CI system to detect any potential
regressions. We (at Arm) can quite easily add a few build tests but then
testing the software stack on QEMU is a bit more involved for various
reasons (first instance of QEMU testing, dependencies on OPTEE, UEFI,
...) so this might have to wait for some time.
Okay, will wait for CI testing.
-Sumit
Regards,
Sandrine
--
TF-A mailing list
TF-A@lists.trustedfirmware.org
https://lists.trustedfirmware.org/mailman/listinfo/tf-a
--
TF-A mailing list
TF-A@lists.trustedfirmware.org