[PATCH v5 02/11] tools: mkeficapsule: add firmwware image signing

[AKASHI Takahiro]

On Fri, Nov 05, 2021 at 11:35:00AM +0900, AKASHI Takahiro wrote:
> On Thu, Nov 04, 2021 at 08:02:40PM -0600, Simon Glass wrote:
> > Hi Takahiro,
> > 
> > On Thu, 4 Nov 2021 at 19:04, AKASHI Takahiro <takahiro.akashi at linaro.org> wrote:
> > >
> > > Hi, Simon,
> > >
> > > On Thu, Nov 04, 2021 at 09:11:59AM -0600, Simon Glass wrote:
> > > > Hi Mark,
> > > >
> > > > On Thu, 4 Nov 2021 at 08:31, Mark Kettenis <mark.kettenis at xs4all.nl> wrote:
> > > > >
> > > > > > From: Simon Glass <sjg at chromium.org>
> > > > > > Date: Wed, 3 Nov 2021 20:51:25 -0600
> > > > > >
> > > > > > Hi Mark,
> > > > > >
> > > > > > On Tue, 2 Nov 2021 at 09:13, Mark Kettenis <mark.kettenis at xs4all.nl> wrote:
> > > > > > >
> > > > > > > > From: Simon Glass <sjg at chromium.org>
> > > > > > > > Date: Tue, 2 Nov 2021 08:56:50 -0600
> > > > > > > >
> > > > > > > > Hi Takahiro,
> > > > > > > >
> > > > > > > > > > - can we just build the tool always?
> > > > > > > > >
> > > > > > > > > This is one of my questions.
> > > > > > > > > Why do you want to do so while there are bunch of tools that are
> > > > > > > > > not always built.
> > > > > > > >
> > > > > > > > Because I think all tools should be built always. It is fine if that
> > > > > > > > happens due to CONFIG options but we should try to avoid making it
> > > > > > > > complicated.
> > > > > > >
> > > > > > > Well, unless this patchset fixes things, we can't, because
> > > > > > > mkeficapsule doesn't build on OpenBSD.  I tried looking into it, but I
> > > > > > > can't figure out how this is even supposed to compile as a host tool:
> > > > > > >
> > > > > > >
> > > > > > > In file included from tools/mkeficapsule.c:8:
> > > > > > > In file included from include/malloc.h:369:
> > > > > > > include/linux/string.h:15:24: error: conflicting types for 'strspn'
> > > > > > > extern __kernel_size_t strspn(const char *,const char *);
> > > > > > >                        ^
> > > > > > > /usr/include/string.h:88:9: note: previous declaration is here
> > > > > > > size_t   strspn(const char *, const char *);
> > > > > >
> > > > > > My guess is that linux/string.h should not be included, or perhaps
> > > > > > __kernel_size_t should be defined to size_t.
> > > > > >
> > > > > > I doubt it would take an age to figure out, with a bit of fiddling.
> > > > >
> > > > > Well, I think the problem is quite fundamental.  Indeed I agree that
> > > > > linux/string.h shouldn't be included.  It gets pulled in because the
> > > > > tools include <malloc.h>.  Modern software really shouldn't include
> > > > > that header anymore, and we removed it in OpenBSD some time ago.  But
> > > > > even with that fixed, things break since the same header gets pulled
> > > > > in from <efi.h>.
> > > > >
> > > > > Redefining __kernel_size_t doesn't provide a way out:
> > > > >
> > > > > tools/mkeficapsule.c:23:16: error: typedef redefinition with different types ('size_t' (aka 'unsigned long') vs 'unsigned int')
> > > > > typedef size_t __kernel_size_t;
> > > > >                ^
> > > > > ./arch/arm/include/asm/posix_types.h:37:23: note: previous definition is here
> > > > > typedef unsigned int            __kernel_size_t;
> > > > >                                                ^
> > > > >
> > > > > This is on an amd64 host, so "unsigned int" clearly is the wrong type
> > > > > for size_t.
> > > > >
> > > > > The fundamental problem seems to be that <efi.h> isn't safe to include
> > > > > in a "host" tool because it includes "target" headers that
> > > > > accidentally resolve to "system" headers on Linux systems.
> > > > >
> > > > > Maybe Takahiro or Heinrich have an idea how to fix that?  But in the
> > > > > meantime it would be good if building this tool would remain optional.
> > > >
> > > > Yes let's ask them to fix that as I agree this sounds wrong. We have
> > > > several efi headers so perhaps just need to have the right stuff in
> > > > each.
> > >
> > > As far as I know, you initially introduced efi.h and efi_api.h.
> > > What is your intent to have the two?
> > >
> > > I think that efi_api.h contains definitions and interfaces defined
> > > in UEFI specification for building EFI application/modules, hence
> > > I believe that it should be target-independent. Right?
> > >
> > > But it *includes* efi.h which also contains some definitions
> > > defined in UEFI specification, while efi.h is only for U-Boot as
> > > UEFI application.
> > >
> > > I suspect that is the root cause.
> > 
> > Yes I think you are right.
> > 
> > > Or should we thoroughly use linux headers like "efi/efi.h"
> > > in this tool?
> > 
> > Well either way, we need host builds to not include U-Boot headers.
> 
> Yeah, but there are still lots of host tools which include U-Boot headers.
> In addition, I'm not quite sure whether *generic* efi headers, like
> efi/efi.h, are available across different host OSs.

I looked through linux's efi headers under /usr/include/efi,
but they don't provide enough set of definitions to make mkeficapsule
buildable. Particularly, capsule-related structure definitions are missing.

So modifying U-Boot headers and removing target-dependent coding
would be more practical.
(I don't know yet whether it is feasible or not.)

Or even adding host-tools-local headers would be more optimal.

-Takahiro Akashi

> -Takahiro Akashi
> 
> > 
> > - Simon
> > 
> > >
> > > -Takahiro Akashi
> > >
> > >
> > > > It is OK to have it optional with a CONFIG, but it should be enabled
> > > > by default, otherwise no one will know it is there.
> > > >
> > > > Can we get the OpenBSD environment into CI or is that just too hard?
> > > >
> > > > Regards,
> > > > Simon