[PATCH] efi_loader: FMP cleanups

[AKASHI Takahiro]

On Tue, Jun 15, 2021 at 08:23:35AM +0300, Ilias Apalodimas wrote:
> On Tue, Jun 15, 2021 at 01:44:58PM +0900, AKASHI Takahiro wrote:
> > On Tue, Jun 15, 2021 at 06:55:50AM +0300, Ilias Apalodimas wrote:
> > > Akashi-san,
> > > 
> > > On Tue, Jun 15, 2021 at 10:51:01AM +0900, AKASHI Takahiro wrote:
> > > > Ilias,
> > > > 
> > > > In this patch, you are trying to address a couple of independent
> > > > issues in a single commit.
> > > > Please split.
> > > > (Heinrich doesn't like that.)
> > 
> > Any comment?
> 
> They are fixing the ESRT table generation, while cleaning up what's already in 
> there.  Besides Heinrich can comment himself if he wants them split or not.

They are fixing "different" problems relating ESRT generation.
That is my point.

> 
> > 
> > > > On Mon, Jun 14, 2021 at 06:10:14PM +0300, Ilias Apalodimas wrote:
> > > > > Right now we allow both of the FMPs (RAW and FIT based) to be installed at
> > > > > the same time.  Moreover we only install those if a CapsuleUpdate is
> > > > > requested.  Since we now have an ESRT table, it makes more sense to
> > > > > unconditionally install the FMP, so any userspace applications (e.g fwupd)
> > > > > can make use of them and trigger an update.
> > > > > 
> > > > > While at it clean up the FMP installation as well.  Chapter 23 of the EFI
> > > > > spec (rev 2.9) says:
> > > > > "A specific updatable hardware firmware store must be represented by 
> > > > > exactly one FMP instance".
> > > > > This is not the case for us, since both of our FMP protocols can be
> > > > > installed at the same time and are controlled by a single 'dfu_alt_info'
> > > > > env variable.
> > > > > So make the config option a choice and allow the user to install one
> > > > > of them at any given time.
> > > > 
> > > > I'd like to say nak in some respects:
> > > > - Although I do understand the UEFI requirement that you mentioned above,
> > > >   FIT and RAW FMP drivers can handle *different* firmware even though
> > > >   they share the same dfu_alt_info.
> > > 
> > > How ?
> > 
> > One idea that I can imagine is that we may be able to utilize
> > "update_image_index", which is currently not used effectively,
> > in order to specify which firmware in dfu_alt_info be handled
> > by either FIT FMP or RAW FMP.
> 
> So it's not being used right now, and the fact is they are at the moment doing
> the same thing. And even if it does, no one in his right mind will create a
> platform and say "Hey let me create half of the capsules as raw and the rest
> of them as FIT, it would be fun to watch users struggle".

You misunderstand me.
Because you asked me about an idea about how to fix the issue,
I answered to it. I have never said that the current code does not
have a problem that you mentioned.
So I said:
> > > >   We should not impose unnecessary restriction if we manage to have some
> > > >   workaround to meet the requirement.
          ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

> Is there anything very specific that you can achieve with FIT capsules that
> you can't achieve with RAW ones (or vice versa), that would justify having
> them both present at the same time?

Yes.
We may have different *firmware* for different software components
and different devices. For example,
You have firmare like U-Boot binary and default variable storage
in different partitions.
On the other hand, you have an extra firmware for a particular
peripheral, like PCI device or anything else, which comes
from a 3rd party vendor of the device.
The former may and can be packed into a single binary in FIT format.
The latter can be used in a separate RAW format as the timing of
updating those firmware is likely to be different.

> > 
> > > >   We should not impose unnecessary restriction if we manage to have some
> > > >   workaround to meet the requirement.
> > > 
> > > It's not the updating part only. It's that the .get_image_info also relies on
> > > the same env variable.
> > 
> > The above idea can and should be applied to GetImageInfo implementation
> > at the same time.
> 
> Yes but can you do it with just changing the env variable now? Or you need to
> add more code into the DFU logic?

Those *meta* data for firmware can be declared/specified outside of FMP,
and be referred to by FMP (and/or ESRT). That is what I meant by:
> > > >   (I still believe that the firmware definition for ESRT should exist
> > > >   elsewhere other than in FMP themselves.)


> > 
> > > Specifically in the fwupd case on an RPI4 with the
> > > dfu set at 'dfu_alt_info=u-boot.bin fat 0 1;' although 2 ERSTs entries were 
> > > populated only one was reported.  Probably because this really does give you 
> > > 2 ways of updating the same flash.
> > > 
> > > >   (I still believe that the firmware definition for ESRT should exist
> > > >   elsewhere other than in FMP themselves.)
> > > 
> > > That's a whole different story, and if we have that, then .get_image_info
> > > should change as well instead of using the DFU information.
> > 
> > I don't think so as I mentioned above.
> 
> And I don't see any benefit from storing the same information in 2 completely
> disjoint entities. 

?

> > 
> > > Because right
> > > now we enabled security (or think we have), while allowing users to set an env
> > > variable which is not authenticated, and completely change what the 
> > > firmware reports (or updates).
> > 
> > This is the point that I mentioned earlier in our private chat,
> > and it's a "whole different" story in this context.
> 
> You mentioned that in the context of "can we install the FMPs during the EFI
> init". Since the variable is interpreted at runtime, we definitely can.  I
> looked back at that chat and saw nothing related to the security problems
> we'll create.

You have referred to this issue in the context of security.
So I said that it was a different story.
The issue that you're trying to address in this patch is *NOT* security.

> In any case the problem here is real, but there are sane ways to avoid it.
> 
> > 
> > > We can always add a huge warning saying
> > > something along the lines of "If you really care this should come with a
> > > CONFIG_ENV_IS_NOWHERE and a boot timeout set to -1".
> > > 
> > > The spec is pretty clear and we allow users to *break* it with a config
> > > option. Arguably this is fine, since the code continues to work fine and
> > > you can perform the updates,  but in essence RAW and FITs are used to update
> > > the same medium right now. You can't have a capsule with half it's contents
> > > describing something RAW and the other half being a FIT.  You have a FIT based 
> > > capsule or a RAW based capsule.
> > 
> > See above.
> 
> I still don't get it. 
> The fact is we have a config option, that if the user decides to set in a
> specific way (and that specific way is 99% of the use cases) we'll break the
> EFI spec.

As I said above, I have never said that the current implementation does
not break EFI spec if not properly used.
So I suggested a possible solution in the previous email as you asked me.

> So unless we add code into the dfu logic, parsing dfu_alt_info and 
> figuring out if the user is allowed to do that or not, I really think those two
> must be treated as mutually exclusive.

I don't think that we need to modify DFU code.

-Takahiro Akashi

> > 
> > > > - We should allow users to add their own FMP drivers and so not call
> > > >   [arch_]efi_load_capsule_drivers() unconditionally
> > > >   even if you don't like "__weak" attribute.
> > > 
> > > I am fine with the __weak attribute.  On the other hand I consider the
> > > current code the defacto way users would use to update their firmware. That's
> > > why I removed the __weak attribute.  If a hardware vendor was to update
> > > their special PCI option ROM or a flash that lives on the secure world they
> > > should install their FMPs on a different handle and leave the current code
> > > as is.
> > 
> > And we should provide an option that disables these existing handle.
> 
> The existing one is not enough?
> 
> > 
> > > > - Selecting only one of FIT and RAW FMPs in sandbox*_defconfig will
> > > >   leave some test cases in pytest skipped.
> > > 
> > > Yea that's unfortunate, but maybe we can just add an extra config on the
> > > sandbox?
> > 
> > Please add another patch that is missing.
> 
> 
> > 
> > -Takahiro Akashi
> >