Hi,
Da: U-Boot u-boot-bounces@lists.denx.de Per conto di Sean Anderson Inviato: venerdì 23 dicembre 2022 00:06
On 11/18/22 15:50, Simon Glass wrote:
Hi Sean,
On Thu, 13 Oct 2022 at 09:41, Sean Anderson sean.anderson@seco.com wrote:
On 10/13/22 3:14 AM, Rasmus Villemoes wrote:
On 12/10/2022 18.28, Sean Anderson wrote:
On 10/12/22 08:59, Simon Glass wrote:
Hi Sean,
On Tue, 11 Oct 2022 at 17:25, Sean Anderson sean.anderson@seco.com wrote: > > When reading data from a FIT image, we must verify the > configuration we get it from. This is because when we have a key > with required = "conf", the image does not need any particular > signature or hash. The configuration is the only required verification, so > we must verify it. > > Users of fit_get_data_node are liable to load unsigned data > unless the user has set required = "image". Even then, they are > vulnerable to mix-and-match attacks. This also affects other > callers of fit_image_verify which don't first call > fit_config_verify, such as source and imxtract. I don't think > there is a backwards-compatible way to fix these interfaces. > Fundamentally, selecting data by image when images are not required > to be verified is unsafe. > > Fixes: 37feaf2f727 ("image: fit: Add some helpers for getting > data") > Signed-off-by: Sean Anderson sean.anderson@seco.com > --- > > boot/image-fit.c | 9 ++++++++- > 1 file changed, 8 insertions(+), 1 deletion(-) > > diff --git a/boot/image-fit.c b/boot/image-fit.c index > 9c04ff78a15..632fd405e29 100644 > --- a/boot/image-fit.c > +++ b/boot/image-fit.c > @@ -1948,7 +1948,14 @@ int fit_get_data_node(const void *fit, > const char *image_uname, > int fit_get_data_conf_prop(const void *fit, const char *prop_name, > const void **data, size_t *size) > { > - int noffset = fit_conf_get_node(fit, NULL); > + int ret, noffset = fit_conf_get_node(fit, NULL); > + > + if (noffset < 0) > + return noffset; > + > + ret = fit_config_verify(fit, noffset); > + if (ret) > + return ret; > > noffset = fit_conf_get_prop_node(fit, noffset, prop_name); > return fit_get_data_tail(fit, noffset, data, size); > -- > 2.35.1.1320.gc452695387.dirty >
This is supposed to work by first verifying the configuration with fit_config_verify(). After that, images in that configuration can be freely loaded, verified by the hash that each image has.
Well, this function was made to replaces several cases where code loaded a FIT image from somewhere, and then wanted to get data from an image based on the configuration. Typically they only want to extract one image, which is the common case for e.g. loading firmware. This idea of this function is to make the common case of "find me the image data from the default config and verify it" easier. If you look at the existing code which uses this function, they do not verify the configuration first. This is mainly because the original versions of this code which I replaced with this function did not verify the configuration either.
So while the above process works for an integrated verification process, like what is done by bootm, it doesn't really work for one-off loading of image data. In particular, the requirements to make this secure (using required = "image" for your key), are not default. When I was trying to determine whether the source command would be OK to use to load some configuration, I looked at it and saw that it did fit_image_verify. I thought that was fine, but if you use required = "config", then all that does is verify the hash.
Yeah, so I've raised this problem with the "source" shell command previously, but never got a satisfactory answer:
https://lore.kernel.org/u-boot/7d711133-d513-5bcb-52f2-a9dbaa9eeded@prevas.d...
So does your patch now mean that it's possible to get a bootscript-wrapped-in-a-FIT-image verified, possibly by adding some dummy (or not so dummy?) "configurations" node? Can you give a complete .its showing how I can build a verifiable boot script?
No. I didn't convert source because it also checks to ensure that the image type is correct, which fit_get_data_node doesn't check. However, it still uses the image name to determine the data to source, which has all the problems as discussed above.
I think to do this right we would need either
- A version of fit_image_verify which treats required = "config" as required = "image". This could be used for cases where the caller doesn't verify a config (such as in cases when the user specifies an image directly).
Without config verification we are subject to mix-and-match attacks.
We already have several functions which just load data from one image (e.g. firmware). There is nothing to mix or match. The load address/entry point are not used, so they do not need to be protected.
Would such modified fit_image_verify require signatures for all the images in the FIT? Or just for those not in a configuration node? As you stated, this would not be backwards-compatible, so why not to define a new required mode, different from "image" and "config", to select this new behaviour?
IMO, a cleaner design should - when required = "config" is set - reject to load data from any image which is not part of a signed and verified configuration. Of course, this behaviour would not be backwards-compatible, too. Therefore, it should be probably be addressed with a "config+" or "config-strict" or whatever else new required mode.
I feel it cleaner as the definitions of the configs in the FIT will state, themselves, what will be accessible/loadable and what will not, once a configuration has been selected. It requires just to hash all images and sign configurations, and it is safe against mix-and-match attacks.
Regards, Massimo
These functions are great candidates for forcing image verification, no config needed.
--Sean
- Add support for specifying a config node. This would be something like the addr#config syntax used by bootm. Of course, this doesn't address existing users of fit_get_data_node.
Yes, let's do this one.
That said, if we do determine the image based on a config, we should definitely verify it.
Yes
Regards, Simon