imgtool: Support for SHA512 for ED25519 usage #2048

de-nordic · 2024-08-23T13:32:43Z

imgtool changes to support the SHA512 and proper ED25519 signare.

This PR is alternative to #2029 where new TLV is introduced.
While working on image changes I have decided that it is not needed to provide new TLV for signature, because:

the SHA512 is newly support hash and there is no logic for it yet so no possible mismatch of SHA and key
it is enough to decide by key and hash TLV whether the ED25519 is double or single hashing.

Depends on #1967

de-nordic · 2024-08-28T10:47:14Z

Found a problem with signature. The problem is as follows:

the imgtool with ed25519 uses PureEdDSA taking sha256 of image as input for signature;
the correct way would be to use directly put image through PureEdDSA and get sha512 of the images signed this way - that is how the PR is doing now;
the problem on MCUboot side is that it is not possible to put image directly through PureEdDSA, which requires loading entire message to memory, so sha512 is calculated as image is read and that is put in HashEdDSA.
PureEdDSA signature and HashEdDSA signature on the same binary string, with the same sha512 over the binary used in second case are not the same.

taltenbach · 2024-08-28T13:55:30Z

@de-nordic I'm looking into this PR because I would like to have proper Ed25519 usage on one of the projects I'm currently working on. To me PureEdDSA should be favored if possible because of the higher collision resistance it provides compared to HashEdDSA. Wouldn't it be possible to use PureEdDSA by modifying this line:

mcuboot/ext/fiat/src/curve25519.c

Line 1110 in 52e2afc

ret = mbedtls_sha512_update_ret(&ctx, message, message_len);

to something that would gradually read and hash the firmware image from flash memory, similarly to what is currently done by bootutil_img_hash? That would make possible to use PureEdDSA since it wouldn't be necessary anymore to load the entire image to RAM.

de-nordic · 2024-08-29T14:54:44Z

@de-nordic I'm looking into this PR because I would like to have proper Ed25519 usage on one of the projects I'm currently working on. To me PureEdDSA should be favored if possible because of the higher collision resistance it provides compared to HashEdDSA. Wouldn't it be possible to use PureEdDSA by modifying this line:

mcuboot/ext/fiat/src/curve25519.c

Line 1110 in 52e2afc

ret = mbedtls_sha512_update_ret(&ctx, message, message_len);

to something that would gradually read and hash the firmware image from flash memory, similarly to what is currently done by bootutil_img_hash? That would make possible to use PureEdDSA since it wouldn't be necessary anymore to load the entire image to RAM.

The PureEdDSA is supposed to be used for signature of entire message at once. Using hash of image as message in PureEdDSA dos not help to improve security.
There are only two ways to do the PureEdDSA in MCUboot:

run PureEdDSA signature verification directly on image, if it is possible to address flash or storage as RAM (pass it as buffer and size to verification function);
divide image into chunks and have separate signature for each chunk (for example one signature every 4k of image).

What we have now is PureEdDSA(sha256(image)), what this PR was supposed to provide PureEdDSA on image and sha512, without introducing new TLV for signature, but it seems that, at this point, I have to do PureEdDSA on sha512, as this allows us, at the cost of lowered collision resistance and security, in respect to proper PureEdDSA, ability to easily support external images for which PureEdDSA may not always be possible; the PureEdDSA on external storage may not be possible if external storage can not be mapped to RAM, where external slot image could be verified at once, or there is not enough memory to just copy the image to RAM and validate it there.

to something that would gradually read and hash the firmware image from flash memory, similarly to what is currently done by bootutil_img_hash? That would make possible to use PureEdDSA since it wouldn't be necessary anymore to load the entire image to RAM.

This is not possible. The HashEdDSA has been provided for that purpose. The PureEdDSA is supposed to work on entire message at once.

nvlsianpu

imgtool part: LGTM
I think ed25519 signature & its hash selection consequences need to be documented (which i can do myself)
You must add your signed-off-by line to the 1st commmit

de-nordic · 2024-08-29T16:19:10Z

imgtool part: LGTM I think ed25519 signature & its hash selection consequences need to be documented (which i can do myself) You must add yours signed-off-by line to the 1st commmit

I have moved back to, as discussed off-line, to doing PureEdDSA on hash of image.
I will follow this PR with update that also does the proper PureEdDSA signature but that will require new TLV and change in MCUboot, in image validation - not only in algorithm but also in how image is passed for verication,

nvlsianpu

LGTM

taltenbach

I'm currently integrating your changes in my project (I cannot wait for this PR to be merged unfortunately), I just noticed two minor things you might want to fix :)

taltenbach · 2024-09-04T21:49:01Z

scripts/imgtool/main.py

@@ -401,6 +402,9 @@ def convert(self, value, param, ctx):
 @click.option('--sig-out', metavar='filename',
              help='Path to the file to which signature will be written. '
              'The image signature will be encoded as base64 formatted string')
+@click.option('--sha', 'user_sha', type=click.Choice(valid_sha), default='auto',
+              help='selected sha algorithm to use; defaults to "auto" which is 256 if '
+              'no cryptographic sygnature is used, or default for signature type')


typo: "signature"

taltenbach · 2024-09-04T21:50:06Z

scripts/imgtool/image.py

@@ -673,14 +740,17 @@ def verify(imgfile, key):
        prot_tlv_size = tlv_off
        hash_region = b[:prot_tlv_size]
        digest = None
+        # Safe bet on sha256, the right one will be identified from TLV
+        digest_sha = '256'


nitpick: digest_sha doesn't seem to be used

de-nordic · 2024-09-05T13:07:17Z

I'm currently integrating your changes in my project (I cannot wait for this PR to be merged unfortunately), I just noticed two minor things you might want to fix :)

Note that there is still missing the ED25519 over sha512 configuration for MCUboot.
There is patch that adds sha512 with PSA crypto, but there is no ED25519 verify that can use it.
Nordic is working on PSA support for ed25519, here https://github.com/nrfconnect/sdk-mcuboot/pull/323/files, but it may take a moment before it gets to upstream.

taltenbach · 2024-09-06T12:18:07Z

Note that there is still missing the ED25519 over sha512 configuration for MCUboot. There is patch that adds sha512 with PSA crypto, but there is no ED25519 verify that can use it. Nordic is working on PSA support for ed25519, here https://github.com/nrfconnect/sdk-mcuboot/pull/323/files, but it may take a moment before it gets to upstream.

Thanks for the details! The project I'm working on is using mbedTLS, not PSA, so I just had to modify a bit sha.h to be able to use SHA-512 with mbedTLS. That's a quick and straightforward change but in case it saves you some time, I can create a PR.

d3zd3z · 2024-09-09T22:06:05Z

the problem on MCUboot side is that it is not possible to put image directly through PureEdDSA, which requires loading entire message to memory, so sha512 is calculated as image is read and that is put in HashEdDSA.

So, right here, you've summarized the entirety of the problem I have with EdDSA. The thing is, I have no problem with the algorithm itself, and as far as I can tell it is sound. But it seems to every implementation seems to have been done by someone who has actually never even tried to use it.

There is absolutely no reason that an implementation of PureEdDSA would ever need to have the entire image in memory. But, as part of the hash, it does require hashing of the image multiple times, so can't just use a simplistic API.

Given that we are basically stuck with the implementations that exist, however, using HashEdDSA seems perfectly fine. HashEdDSA basically reduces the security of EdDSA to the same as ECDSA, and, for that matter, every other signature algorithm.

d3zd3z

A little bit about the commit text:

I'd remove the "not a perfect usage". Using EdDSA to sign the hash of the image is perfectly acceptable, even described in the original documentation for the algorithm.
Change "sing" to "sign", unless perhaps we've become musical in nature.
I wouldn't really bother saying that larger hashes bring higher collision resistance. Unless we plan on doing something like signing all of the atoms in universe, individually, even SHA256 will never encounter a collision. Larger hashes are generally more about having more information to give to a large signature size.

scripts/imgtool/image.py

de-nordic · 2024-09-10T13:35:36Z

* I'd remove the "not a perfect usage". Using EdDSA to sign the hash of the image is perfectly acceptable, even described in the original documentation for the algorithm.

OK, will fix the comment.

* Change "sing" to "sign", unless perhaps we've become musical in nature.

Heh, no.

* I wouldn't really bother saying that larger hashes bring higher collision resistance. Unless we plan on doing something like signing all of the atoms in universe, individually, even SHA256 will never encounter a collision. Larger hashes are generally more about having more information to give to a large signature size.

OK, will fix. Still some of implementations, like PSA ed25519, require sha512 otherwise they refuse to work.

de-nordic · 2024-09-10T14:38:58Z

So, right here, you've summarized the entirety of the problem I have with EdDSA. The thing is, I have no problem with the algorithm itself, and as far as I can tell it is sound. But it seems to every implementation seems to have been done by someone who has actually never even tried to use it.

I think the general idea is to not let anyone feed you whatever they want as something you try to verify signature on. If you have some black-box crypto device you can just feed it message and signature and get verification for it, without pulling hash out of the black box. If you can provide hash, same may be possible for an attacker (?), which means that having valid signature-hash pair is enough to validate any message.

There is absolutely no reason that an implementation of PureEdDSA would ever need to have the entire image in memory. But, as part of the hash, it does require hashing of the image multiple times, so can't just use a simplistic API.

Above would be the reason, although if an implementation would just have callback "give-me-chunk" then that would be enough to pass whatever the implementation wants.

Given that we are basically stuck with the implementations that exist, however, using HashEdDSA seems perfectly fine. HashEdDSA basically reduces the security of EdDSA to the same as ECDSA, and, for that matter, every other signature algorithm.

We do not have HashEdDSA, we have SHA256-ED25519-SHA512 or SHA512-ED25519-SHA512, as far as I understand http://ed25519.cr.yp.to/eddsa-20150704.pdf and https://www.rfc-editor.org/rfc/rfc8032.html#page-9

de-nordic · 2024-09-11T13:49:55Z

Good point, I stand corrected. Do you know any ed25519 implementation, in any library/language, that is not built with sha-512 as the hash?

No. Still on the loop regarding the shaxyz-ed25519-sha512 naming? It is just a designation, and a correct one.

Which section mentions this? Is there a rationale or do they just pick sha-512 because it's already there as the hash?

Actually they just state (in http://ed25519.cr.yp.to/eddsa-20150704.pdf) that you can use any hash function as input to the PureEdDSA, which kinda makes sense - it just signs/verifies message, it does not really know what it is.
For rationale, ask the authors of both documents - and if they say "yes", then what?

Using what you have is still a good rationale, considering the fact that you only need to bring in implementation of a single hash function; also if there is hardware support there may be sha512 accelerator, because there is ed25519 hardware support, it does not mean you will have other sha implementations - which again means that you have to bring either driver or software support for other sha.
The HashEdDSA expects sha512 by algorithm definition; the PSA implementation of HashEdDSA requires SHA512 and will refuse to work on sha256. I do not see why we should not then use sha512 as message for PureEdDSA in the sha-512-ed25519 scheme. Even if that does not improve security, as you state, the reduction in required support for two sha algs is already an improvement.

Unrelated point but does mcumgr already support testing an image whose hash is sha-512?

Not yet, but I work on the implementation, with use of the PSA of PureEdDSA and the PureEdDSA(sha512) over image and have been using the imgtool here to generate images for tests.

utzig · 2024-09-11T14:06:12Z

Good point, I stand corrected. Do you know any ed25519 implementation, in any library/language, that is not built with sha-512 as the hash?

No. Still on the loop regarding the shaxyz-ed25519-sha512 naming? It is just a designation, and a correct one.

Which section mentions this? Is there a rationale or do they just pick sha-512 because it's already there as the hash?

Actually they just state (in http://ed25519.cr.yp.to/eddsa-20150704.pdf) that you can use any hash function as input to the PureEdDSA, which kinda makes sense - it just signs/verifies message, it does not really know what it is. For rationale, ask the authors of both documents - and if they say "yes", then what?

Using what you have is still a good rationale, considering the fact that you only need to bring in implementation of a single hash function; also if there is hardware support there may be sha512 accelerator, because there is ed25519 hardware support, it does not mean you will have other sha implementations - which again means that you have to bring either driver or software support for other sha. The HashEdDSA expects sha512 by algorithm definition; the PSA implementation of HashEdDSA requires SHA512 and will refuse to work on sha256. I do not see why we should not then use sha512 as message for PureEdDSA in the sha-512-ed25519 scheme. Even if that does not improve security, as you state, the reduction in required support for two sha algs is already an improvement.

Unrelated point but does mcumgr already support testing an image whose hash is sha-512?

Not yet, but I work on the implementation, with use of the PSA of PureEdDSA and the PureEdDSA(sha512) over image and have been using the imgtool here to generate images for tests.

Sorry for asking.

de-nordic · 2024-09-11T14:36:38Z

Sorry for asking.

No problem.

de-nordic · 2024-09-18T11:10:29Z

@d3zd3z I have addressed your comments, can you re-review?

taltenbach · 2024-09-19T11:55:10Z

LGTM, using #2066 I was able to test the images generated by imgtool.py on an STM32F4 MCU. Works fine 👍

The adds support for hashing image with SHA512, to allow SHA512-ED25519-SHA512 signature. To support above --sha parameter has been added that can take value: auto, 256, 384, 512 to select sha, where auto brings the default behaviour, or current, behaviour. The sha provided here is tested against key so not all combinations are supported. Upstream PR: mcu-tools/mcuboot#2048 Signed-off-by: Dominik Ermel <dominik.ermel@nordicsemi.no>

The adds support for hashing image with SHA512, to allow SHA512-ED25519-SHA512 signature. To support above --sha parameter has been added that can take value: auto, 256, 384, 512 to select sha, where auto brings the default behaviour, or current, behaviour. The sha provided here is tested against key so not all combinations are supported. Upstream PR: mcu-tools/mcuboot#2048 Signed-off-by: Dominik Ermel <dominik.ermel@nordicsemi.no> (cherry picked from commit 3a28585)

de-nordic requested review from nvlsianpu, davidvincze, d3zd3z and michalek-no August 23, 2024 13:32

de-nordic force-pushed the ed25519_single_sha branch 3 times, most recently from 332ac85 to cee234c Compare August 23, 2024 15:46

de-nordic added the [DNM] Do Not Merge label Aug 28, 2024

de-nordic mentioned this pull request Aug 29, 2024

RFC: bootutil: Add TLV for Pure (over image) signature #2029

Merged

nvlsianpu reviewed Aug 29, 2024

View reviewed changes

de-nordic changed the title ~~Support for SHA512 and proper ED25519 usage~~ Support for SHA512 and for ED25519 usage Aug 29, 2024

de-nordic requested a review from nvlsianpu August 29, 2024 16:17

nvlsianpu reviewed Aug 29, 2024

View reviewed changes

de-nordic force-pushed the ed25519_single_sha branch from cee234c to dd40e1b Compare August 29, 2024 17:50

de-nordic mentioned this pull request Sep 3, 2024

boot: SHA512 verification #1967

Merged

de-nordic force-pushed the ed25519_single_sha branch 2 times, most recently from 3d23b73 to 8261a11 Compare September 5, 2024 12:01

taltenbach reviewed Sep 5, 2024

View reviewed changes

de-nordic force-pushed the ed25519_single_sha branch from 8261a11 to 0d42481 Compare September 5, 2024 13:01

de-nordic removed the [DNM] Do Not Merge label Sep 5, 2024

d3zd3z requested changes Sep 9, 2024

View reviewed changes

scripts/imgtool/image.py Show resolved Hide resolved

de-nordic changed the title ~~Support for SHA512 and for ED25519 usage~~ Support for SHA512 for ED25519 usage Sep 11, 2024

de-nordic mentioned this pull request Sep 13, 2024

imgtool: Add --pure switch to allow calculating signatures over entire images #2063

Open

de-nordic changed the title ~~Support for SHA512 for ED25519 usage~~ imgtool: Support for SHA512 for ED25519 usage Sep 13, 2024

nvlsianpu approved these changes Sep 18, 2024

View reviewed changes

de-nordic requested a review from taltenbach September 19, 2024 11:32

taltenbach mentioned this pull request Sep 19, 2024

bootutil: Add support for Ed25519 using SHA-512 with mbedTLS #2066

Merged

taltenbach approved these changes Sep 19, 2024

View reviewed changes

nordicjm approved these changes Sep 27, 2024

View reviewed changes

d3zd3z approved these changes Oct 1, 2024

View reviewed changes

nordicjm merged commit c894d04 into mcu-tools:main Oct 1, 2024
58 checks passed

GeorgeCGV mentioned this pull request Oct 28, 2024

imgtool: fix signing for fix-sig-pubkey public rsa #2105

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

imgtool: Support for SHA512 for ED25519 usage #2048

imgtool: Support for SHA512 for ED25519 usage #2048

de-nordic commented Aug 23, 2024 •

edited

Loading

de-nordic commented Aug 28, 2024 •

edited

Loading

taltenbach commented Aug 28, 2024 •

edited

Loading

de-nordic commented Aug 29, 2024

nvlsianpu left a comment •

edited

Loading

de-nordic commented Aug 29, 2024

nvlsianpu left a comment

taltenbach left a comment

taltenbach Sep 4, 2024

taltenbach Sep 4, 2024

de-nordic commented Sep 5, 2024

taltenbach commented Sep 6, 2024

d3zd3z commented Sep 9, 2024

d3zd3z left a comment

de-nordic commented Sep 10, 2024

de-nordic commented Sep 10, 2024

de-nordic commented Sep 11, 2024

utzig commented Sep 11, 2024

de-nordic commented Sep 11, 2024

de-nordic commented Sep 18, 2024

taltenbach commented Sep 19, 2024

imgtool: Support for SHA512 for ED25519 usage #2048

imgtool: Support for SHA512 for ED25519 usage #2048

Conversation

de-nordic commented Aug 23, 2024 • edited Loading

de-nordic commented Aug 28, 2024 • edited Loading

taltenbach commented Aug 28, 2024 • edited Loading

de-nordic commented Aug 29, 2024

nvlsianpu left a comment • edited Loading

Choose a reason for hiding this comment

de-nordic commented Aug 29, 2024

nvlsianpu left a comment

Choose a reason for hiding this comment

taltenbach left a comment

Choose a reason for hiding this comment

taltenbach Sep 4, 2024

Choose a reason for hiding this comment

taltenbach Sep 4, 2024

Choose a reason for hiding this comment

de-nordic commented Sep 5, 2024

taltenbach commented Sep 6, 2024

d3zd3z commented Sep 9, 2024

d3zd3z left a comment

Choose a reason for hiding this comment

de-nordic commented Sep 10, 2024

de-nordic commented Sep 10, 2024

de-nordic commented Sep 11, 2024

utzig commented Sep 11, 2024

de-nordic commented Sep 11, 2024

de-nordic commented Sep 18, 2024

taltenbach commented Sep 19, 2024

de-nordic commented Aug 23, 2024 •

edited

Loading

de-nordic commented Aug 28, 2024 •

edited

Loading

taltenbach commented Aug 28, 2024 •

edited

Loading

nvlsianpu left a comment •

edited

Loading