[PATCH] efi_loader: Measure the loaded DTB

[Heinrich Schuchardt]

Am 8. Dezember 2022 09:01:26 MEZ schrieb Ilias Apalodimas <ilias.apalodimas at linaro.org>:
>Hi Heinrich
>
>On Thu, Dec 08, 2022 at 05:12:26AM +0100, Heinrich Schuchardt wrote:
>> On 12/7/22 16:11, Etienne Carriere wrote:
>> > Measures the DTB passed to the EFI application upon new boolean config
>> > switch CONFIG_EFI_TCG2_PROTOCOL_MEASURE_DTB. For platforms where the
>> > content of the DTB passed to the OS can change across reboots, there is
>> > not point measuring it hence the config switch to allow platform to not
>> > embed this feature.
>> > 
>> > Co-developed-by: Ilias Apalodimas <ilias.apalodimas at linaro.org>
>> > Signed-off-by: Ilias Apalodimas <ilias.apalodimas at linaro.org>
>> > Signed-off-by: Etienne Carriere <etienne.carriere at linaro.org>
>> > ---
>> >   cmd/bootefi.c             |  9 +++++++++
>> >   include/efi_loader.h      |  2 ++
>> >   include/efi_tcg2.h        | 10 ++++++++++
>> >   include/tpm-v2.h          |  2 ++
>> >   lib/efi_loader/Kconfig    | 12 ++++++++++++
>> >   lib/efi_loader/efi_tcg2.c | 36 ++++++++++++++++++++++++++++++++++++
>> >   6 files changed, 71 insertions(+)
>> > 
>> > diff --git a/cmd/bootefi.c b/cmd/bootefi.c
>> > index 2a7d42925d..56e4a1909f 100644
>> > --- a/cmd/bootefi.c
>> > +++ b/cmd/bootefi.c
>> > @@ -315,6 +315,15 @@ efi_status_t efi_install_fdt(void *fdt)
>> >   		return EFI_LOAD_ERROR;
>> >   	}
>> > 
>> > +	/* Measure the installed DTB */
>> > +	if (CONFIG_IS_ENABLED(EFI_TCG2_PROTOCOL_MEASURE_DTB)) {
>> > +		ret = efi_tcg2_measure_dtb(fdt);
>> 
>> Why measure here and not in efi_tcg2_measure_efi_app_invocation() where
>> we measure the SMBIOS table every time when an app is started with
>> StartImage()? Which may mean measuring it multiple times like when
>> starting iPXE, when starting GRUB, and again when starting Linux.
>
>IIRC the spec says you should try and measure things once.  On top of that
>it has no wording for DTB,  but for ACPI tables it says they should be measured prior
>to any fix-ups.  We pretty much copied the recommendations for ACPI apart
>from when we measure the data.  For ACPI it says "ACPI flash data prior to
>any modifications" [0] but since U-Boot allows us to change the loaded DTB we
>tried to measure the one we end up installing in the config table.

If for ACPI it is sufficient to measure before fixups, doing the same for dtbs should be adequate. This would avoid variable parts like random MAC addresses, boot-hart ID, kaslr seed but should include applied overlays. So just measuring the dtb passed into bootefi or the fallback dtb.

Why do we measure SMBIOS multiple times?


>
>> 
>> What is the value of measuring the device-tree here when GRUB or
>> systemd-boot afterwards load a different device-tree or modify the
>> provided device-tree?
>
>If they modify if there's some usage since you, in theory, trust and
>hopefully verified those binaries and the modifications they are about to
>perform.  
>If they end up replacing it there's no point in measuring, but that's exactly why
>this is under a Kconfig option.  If they load a different DTB then they are responsible
>for measuring it?
>
>[...]
>
>> > +			log_err("ERROR: failed to measure DTB\n");
>> > +			return ret;
>> > +		}
>> > +	}
>> > +
>> 
>> %s/using/using the/
>> 
>> > +	  the passed DTB contains data that change across platform reboots
>> > +	  and cannot be used has a predictable measurement. Otherwise
>> 
>> How should the user know this? Shall we create dependencies to other
>> Kconfig symbols? e.g.
>> 
>>     depends on !NET_RANDOM_ETHADDR
>> 
>
>That's not a bad idea.  Maybe we can add the ones that make sense and keep
>extending it if we find more values affect the 'randomness'
>The whole point of having a discrete Kconfig on that is that we don't
>really know what kind of random things people end up dumping on their DTB.
>
>> > +	  this feature allows better measurement of the system boot
>> > +	  sequence.
>> > +
>> >   config EFI_LOAD_FILE2_INITRD
>> >   	bool "EFI_FILE_LOAD2_PROTOCOL for Linux initial ramdisk"
>> >   	default y
>> > diff --git a/lib/efi_loader/efi_tcg2.c b/lib/efi_loader/efi_tcg2.c
>> > index a525ebf75b..51c9d80828 100644
>> > --- a/lib/efi_loader/efi_tcg2.c
>> > +++ b/lib/efi_loader/efi_tcg2.c
>> > @@ -2175,6 +2175,42 @@ out1:
>> >   	return ret;
>> >   }
>> > 
>> > +/**
>> > + * efi_tcg2_measure_dtb() - measure the dtb used to boot our OS
>> > + *
>> > + * @fdt: pointer to the device tree blob
>> > + *
>> > + * Return:	status code
>> > + */
>> > +efi_status_t efi_tcg2_measure_dtb(void *fdt)
>> > +{
>> > +	efi_status_t ret;
>> > +	struct uefi_platform_firmware_blob2 *blob;
>> > +	struct udevice *dev;
>> > +	u32 event_size;
>> > +
>> > +	if (!is_tcg2_protocol_installed())
>> > +		return EFI_SUCCESS;
>> > +
>> > +	ret = platform_get_tpm2_device(&dev);
>> > +	if (ret != EFI_SUCCESS)
>> > +		return EFI_SECURITY_VIOLATION;
>> > +
>> > +	event_size = sizeof(*blob) + sizeof(EFI_DTB_EVENT_STRING) + fdt_totalsize(fdt);
>> > +	blob = calloc(1, event_size);
>> > +	if (!blob)
>> > +		return EFI_OUT_OF_RESOURCES;
>> > +
>> > +	blob->blob_description_size = sizeof(EFI_DTB_EVENT_STRING);
>> > +	memcpy(blob->data, EFI_DTB_EVENT_STRING, blob->blob_description_size);
>> > +	memcpy(blob->data + blob->blob_description_size, fdt, fdt_totalsize(fdt));
>> > +
>> > +	ret = tcg2_measure_event(dev, 0, EV_POST_CODE, event_size, (u8 *)blob);
>> 
>> We should ignore "free space" surrounding blocks. This free space is
>> expected to contain random data. E.g in copy_fdt() we do not zero out
>> the memory after the strings block.

What are your thoughts on ignoring non-coding areas? This would avoid a lot of randomness.


>> 
>> On the RISC-V architecture we always write the boot-hart ID into the
>> device-tree to be backward compatible to old kernels. Which other
>> device-tree information is random and needs to be ignored when measuring?
>> 
>> Is there a standard defining into which PCR the DTB shall be measured?
>> Why did you choose pcr_index=0?
>
>The TCG spec doesn't have any wording regarding DTB atm.  PCR0 is what's
>defined in the spec for ACPI tables

That could be worth a code comment.

Best regards

Heinrich

>
>[0] https://trustedcomputinggroup.org/wp-content/uploads/TCG_PCClient_PFP_r1p05_v23_pub.pdf (page 29)
>
>> 
>> Why measure device-trees but not ACPI tables?
>
>You can measure those as well and we probably should,  but this patch is
>only trying to address DTBs.
>
>Regards
>/Ilias
>
>> 
>> Best regards
>> 
>> Heinrich
>> 
>> > +
>> > +	free(blob);
>> > +	return ret;
>> > +}
>> > +
>> >   /**
>> >    * efi_tcg2_measure_efi_app_invocation() - measure efi app invocation
>> >    *
>>