Re: [PATCH V4 0/3] scripts: dtc: Build fdtoverlay

[David Gibson]

On Thu, Feb 04, 2021 at 08:25:23AM -0600, Rob Herring wrote:
> On Sun, Jan 31, 2021 at 10:39 PM David Gibson
> <david@xxxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Mon, Jan 25, 2021 at 09:42:21PM -0600, Frank Rowand wrote:
> > > Hi David,
> > >
> > > On 1/22/21 12:34 AM, David Gibson wrote:
> > > > On Wed, Jan 20, 2021 at 10:47:40AM +0530, Viresh Kumar wrote:
> > > >> +David.
> > > >>
> > > >> On 19-01-21, 11:12, Frank Rowand wrote:
> > > >>> On 1/12/21 2:28 AM, Viresh Kumar wrote:
> > > >>>> We will start building overlays for platforms soon in the kernel and
> > > >>>> would need fdtoverlay tool going forward. Lets start fetching and
> > > >>>> building it.
> > > >>>>
> > > >>>> While at it, also remove fdtdump.c file, which isn't used by the kernel.
> > > >>>>
> > > >>>> V4:
> > > >>>> - Don't fetch and build fdtdump.c
> > > >>>> - Remove fdtdump.c
> > > >>>>
> > > >>>> Viresh Kumar (3):
> > > >>>>   scripts: dtc: Add fdtoverlay.c to DTC_SOURCE
> > > >>>>   scripts: dtc: Build fdtoverlay tool
> > > >>>>   scripts: dtc: Remove the unused fdtdump.c file
> > > >>>>
> > > >>>>  scripts/dtc/Makefile             |   6 +-
> > > >>>>  scripts/dtc/fdtdump.c            | 163 -------------------------------
> > > >>>>  scripts/dtc/update-dtc-source.sh |   6 +-
> > > >>>>  3 files changed, 8 insertions(+), 167 deletions(-)
> > > >>>>  delete mode 100644 scripts/dtc/fdtdump.c
> > > >>>>
> > > >>>
> > > >>> My first inclination was to accept fdtoverlay, as is, from the upstream
> > > >>> project.
> > > >>>
> > > >>> But my experiences debugging use of fdtoverlay against the existing
> > > >>> unittest overlay files has me very wary of accepting fdtoverlay in
> > > >>> it's current form.
> > > >>>
> > > >>> As an exmple, adding an overlay that fails to reply results in the
> > > >>> following build messages:
> > > >>>
> > > >>>    linux--5.11-rc> make zImage
> > > >>>    make[1]: Entering directory '/local/frowand_nobackup/src/git_linus/build/dragon_linus_5.11-rc'
> > > >>>      GEN     Makefile
> > > >>>      CALL    /local/frowand_nobackup/src/git_linus/linux--5.11-rc/scripts/checksyscalls.sh
> > > >>>      CALL    /local/frowand_nobackup/src/git_linus/linux--5.11-rc/scripts/atomic/check-atomics.sh
> > > >>>      CHK     include/generated/compile.h
> > > >>>      FDTOVERLAY drivers/of/unittest-data/static_test.dtb
> > > >>>
> > > >>>    Failed to apply 'drivers/of/unittest-data/overlay.dtb': FDT_ERR_NOTFOUND
> > > >>>    make[4]: *** [/local/frowand_nobackup/src/git_linus/linux--5.11-rc/drivers/of/unittest-data/Makefile:96: drivers/of/unittest-data/static_test.dtb] Error 1
> > > >>>    make[3]: *** [/local/frowand_nobackup/src/git_linus/linux--5.11-rc/scripts/Makefile.build:496: drivers/of/unittest-data] Error 2
> > > >>>    make[2]: *** [/local/frowand_nobackup/src/git_linus/linux--5.11-rc/scripts/Makefile.build:496: drivers/of] Error 2
> > > >>>    make[1]: *** [/local/frowand_nobackup/src/git_linus/linux--5.11-rc/Makefile:1805: drivers] Error 2
> > > >>>    make[1]: Leaving directory '/local/frowand_nobackup/src/git_linus/build/dragon_linus_5.11-rc'
> > > >>>    make: *** [Makefile:185: __sub-make] Error 2
> > > >>>
> > > >>>
> > > >>> The specific error message (copied from above) is:
> > > >>>
> > > >>>    Failed to apply 'drivers/of/unittest-data/overlay.dtb': FDT_ERR_NOTFOUND
> > > >>>
> > > >>> which is cryptic and does not even point to the location in the overlay that
> > > >>> is problematic.  If you look at the source of fdtoverlay / libfdt, you will
> > > >>> find that FDT_ERR_NOTFOUND may be generated in one of many places.
> > > >>>
> > > >>> I do _not_ want to do a full review of fdtoverlay, but I think that it is
> > > >>> reasonable to request enhancing fdtoverlay in the parent project to generate
> > > >>> usable error messages before enabling fdtoverlay in the Linux kernel tree.
> > > >
> > >
> > > > That's... actually much harder than it sounds.  fdtoverlay is
> > > > basically a trivial wrapper around the fdt_overlay_apply() function in
> > > > libfdt.  Matching the conventions of the rest of the library, really
> > > > it's only way to report errors is a single error code.
> > > >
> > > > Returning richer errors is not an easy problem in a C library,
> > > > especially one designed to be usable in embedded systems, without an
> > > > allocator or much else available.
> > > >
> > > > Of course it would be possible to write a friendly command line tool
> > > > specifically for applying overlays, which could give better errors.
> > > > fdtoverlay as it stands isn't really that - it was pretty much written
> > > > just to invoke fdt_overlay_apply() in testcases.
> > >
> > > Thank you for providing that context.
> > >
> > > I do not know if there is a way to enable the code that is currently in libfdt
> > > to both be useful as an embedded library (for example, U-boot seems to often
> > > have a need to keep memory usage very small) and also be part of a tool with
> > > effective warning and error messages.
> >
> > Yeah, I don't know either.
> >
> > > Before having looked at libfdt only at a cursory level while debugging the proposed
> > > use of fdtoverlay in Linux, my first thought was that maybe it would be possible
> > > to add warning and error messages within "#ifdef" blocks, or other ways that
> > > cause the error code to _not_ be compiled as part of library version of libfdt,
> > > but only be compiled as part of fdtoverlay _when built in the Linux kernel_
> > > (noting that the proposed Linux patch builds the libfdt files as part of
> > > the fdtoverlay compile instead of as a discrete library).  After looking at
> > > the libfdt source a tiny bit more carefully, I would probably shoot down this
> > > suggestion, as it makes the source code uglier and harder to understand and
> > > maintain for the primary purpose of being an embedded library.
> >
> > Oof.  That sounds really ugly, but maybe it could be pulled off.
> >
> > > Do you have any thoughts on how warning and error messages could be added,
> > > or if it is even possible?  Or maybe your suggestion of writing a "friendly
> > > command line tool specifically for applying overlays" is the path that
> > > Viresh should pursue?
> >
> > I think at this stage it's a matter of trying a few approaches and
> > seeing what works out.
> 
> Another way would be applying overlays to dtc's live tree. This could
> apply overlays from dts in addition to dtb. It could be a plug-in if
> we ever get that finished up.

This is actually a really interesting idea, because in a sense dtc
already *does* apply overlays.  It's just that it effectively resolves
as it is parsing, rather than realizing separate overlay objects then
merging as a separate step.

I would actually like to change that, so that it *does* explicitly
produce a chain of overlays internally.  That has advantages for the
checking code, because some checks make sense to apply to individual
overlay fragments, but some only make sense on a fully resolved tree.

As a bonus, it could handle this use case.  Unlike libfdt, dtc is a
much more normal userspace program and adding extra verbose debugging
is no realy problem.

It probably is more work in the short term, though.

> The downside of this is not testing libfdt's code and possible
> differences between 2 implementations.

That can be mitigated by having a bunch of examples in the testsuite
where we cross compare fdtoverlay's output with dtc's.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

Attachment: signature.asc
Description: PGP signature