[libcamera-devel] [oe] [meta-multimedia][PATCH] libcamera: fix packaging and installation

[Laurent Pinchart]

Hi Andrey,

On Mon, Jul 27, 2020 at 07:17:04PM +0300, Andrey Konovalov wrote:
> On 27.07.2020 19:03, Laurent Pinchart wrote:
> > On Mon, Jul 27, 2020 at 06:46:47PM +0300, Andrey Konovalov wrote:
> >> On 27.07.2020 14:11, Laurent Pinchart wrote:
> >>> On Mon, Jul 27, 2020 at 12:58:23PM +0300, Andrey Konovalov wrote:
> >>>> On 27.07.2020 12:42, Kieran Bingham wrote:
> >>>>> On 27/07/2020 10:21, Andrey Konovalov wrote:
> >>>>>> libcamera checks if RPATH or RUNPATH dynamic tag is present in
> >>>>>> libcamera.so. If it does, it assumes that libcamera binaries are
> >>>>>> run directly from the build directory without installing them, and
> >>>>>> tries to use resorces like IPA modules from the build directory.
> >>>>>> Mainline meson strips RPATH/RUNPATH out at install time (for
> >>>>>> meson versions up to 0.54; the things are somewhat changed in 0.55).
> >>>>>> But openembedded-core patches meson to disable RPATH/RUNPATH removal.
> >>>>>> That's why we need to remove this tag manually in do_install_append().
> >>>>>
> >>>>> Uh oh, what's changed... (I'll have to go take a look).
> >>>>>
> >>>>>     -
> >>>>> https://mesonbuild.com/Release-notes-for-0-55-0.html#rpath-removal-now-more-careful
> >>>>>
> >>>>> If we're reliant upon meson behaviour which is no longer consistent,
> >>>>> then we are going to have to do something else in libcamera.
> >>>>
> >>>> I haven't tried meson 0.55 yet, but my impression was that 0.55 should work
> >>>> just as before for "usual" (as per libcamera's README) libcamera build. And
> >>>> starting from 0.55 the patch in openembedded-core to disable RPATH/RUNPATH removal
> >>>> *might* be dropped - if all the packages would be able to set RUNPATH to
> >>>> what they need, and meson would detect that OK in all those cases.
> >>>
> >>> I think that if the problem is caused by a meson patch in openembedded,
> >>> then it would make sense to fix it there. We can decide to address the
> >>> issue in libcamera itself if it's found to affect other distributions
> >>> too, or if meson's behaviour changes in an incompatible way.
> >>
> >> It looks like it is not openembedded only issue:
> >>
> >> -------- Forwarded Message --------
> >> Subject: [libcamera-devel] [PATCH v4 0/2] package/libcamera: bump version to 96fab38
> >> Date: Tue, 16 Jun 2020 20:59:49 +0200
> >> From: Peter Seiderer <ps.report at gmx.net>
> >> To: buildroot at busybox.net
> >> CC: libcamera-devel at lists.libcamera.org, Yann E . MORIN <yann.morin.1998 at free.fr>
> >>
> >> <snip>
> >>
> >> With the following patch libcamera is forced to believe it is running
> >> in a installed environment:
> >>
> >> diff --git a/src/libcamera/utils.cpp b/src/libcamera/utils.cpp
> >> index d55338f..4ff9dac 100644
> >> --- a/src/libcamera/utils.cpp
> >> +++ b/src/libcamera/utils.cpp
> >> @@ -346,15 +346,18 @@ details::StringSplitter split(const std::string &str, const std::string &delim)
> >>     */
> >>    bool isLibcameraInstalled()
> >>    {
> >> +#if 0
> >>    	/*
> >>    	 * DT_RUNPATH (DT_RPATH when the linker uses old dtags) is removed on
> >>    	 * install.
> >>    	 */
> >>    	for (const ElfW(Dyn) *dyn = _DYNAMIC; dyn->d_tag != DT_NULL; ++dyn) {
> >> -		if (dyn->d_tag == DT_RUNPATH || dyn->d_tag == DT_RPATH)
> >> +		if (dyn->d_tag == DT_RUNPATH || dyn->d_tag == DT_RPATH) {
> >> +			printf("XXXXX - dyn->d_un.d_ptr: %s\n", (char*)dyn->d_un.d_ptr);
> >>    			return false;
> >> +		}
> >>    	}
> >> -
> >> +#endif
> >>    	return true;
> >>    }
> >>
> >> Maybe this is because of the buildroot local meson patch ([1]), leading
> >> to an empty (but not absent) RPATH?
> > 
> > buildroot preserves empty RPATH when installing. Maybe we could adapt
> > isLibcameraInstalled() to return true only if RPATH is found *and* not
> > empty ?
> > 
> > For openembedded, why is RPATH stripping skipped ?
> 
> Not sure if I can give a proper explanation, so I would better refer you to:
> 
> https://patches.openembedded.org/patch/156584/
> https://github.com/mesonbuild/meson/issues/2567 (rpurdie's comments especially)

Thank you for the pointers, it's pretty interesting.

> Maybe the fix done in meson 0.55 would let distributions to re-enable RPATH stripping:
>    https://mesonbuild.com/Release-notes-for-0-55-0.html#added-ability-to-specify-targets-in-meson-compile
>    https://github.com/mesonbuild/meson/pull/7103
>    https://github.com/mesonbuild/meson/pull/7472
> - but I don't know if/when this could happen. For me it looks like a lot of re-testing
> at least (distros have a whole lot of packages which could be affected).

Yes, it will likely take time, even if I think it's the right solution.
In the meantime, I think we'll have to handle this in the OE and
buildroot recipes, I don't really see what else we could do. It may be
possible to improve the implementation in libcamera by checking the
R(UN)PATH value (it's tricky though, glibc and musl handle this quite
differently, where glibc patches the _DYNAMIC entries with pointers to
the string table in memory, while musl doesn't AFAIR), but it would only
help if we can get the build system to strip the custom R(UN)PATH entry
that we add. If nothing is stripped, I think we're stuck.

> >> <snip>
> >>
> >> [0:02:18.125804232] [252] DEBUG IPAManager ipa_manager.cpp:316 IPA module /usr/lib/libcamera/ipa_rpi.so signature is not valid
> >>
> >> <snip>
> >>
> >> This can be avoided with the following patch/hack (disable signature check):
> >>
> >> diff --git a/src/libcamera/ipa_manager.cpp b/src/libcamera/ipa_manager.cpp
> >> index 505cf61..3d64898 100644
> >> --- a/src/libcamera/ipa_manager.cpp
> >> +++ b/src/libcamera/ipa_manager.cpp
> >> @@ -301,6 +301,9 @@ std::unique_ptr<IPAProxy> IPAManager::createIPA(PipelineHandler *pipe,
> >>
> >>    bool IPAManager::isSignatureValid(IPAModule *ipa) const
> >>    {
> >> +#if 1
> >> +	return true;
> >> +#else
> >>    #if HAVE_IPA_PUBKEY
> >>    	File file{ ipa->path() };
> >>    	if (!file.open(File::ReadOnly))
> >> @@ -320,6 +323,7 @@ bool IPAManager::isSignatureValid(IPAModule *ipa) const
> >>    #else
> >>    	return false;
> >>    #endif
> >> +#endif
> >>    }
> >>
> >>    } /* namespace libcamera */
> >>
> >>
> >> Maybe related to the buildroot finalize and/or sanitizing RPATH in target tree
> >> step (and/or strip after install with BR2_ENABLE_DEBUG=y/BR2_STRIP_strip=y
> >> enabled)?
> > 
> > For this, let's first see how module re-signing works with
> > openemebedded, possibly improving the resigning script. If it's
> > successful, I think we can then use the same methods for buildroot and
> > other distributions. A packaging document to explain all this would be
> > useful.
> 
> Sounds good.
> 
> >> -------- End of Forwarded Message --------
> >>
> >>>>> /me sighs ...
> >>>>>
> >>>>>> IPA module is signed (with openssl dgst) after it is built. But
> >>>>>> during packaging the OE build system 1) splits out debugging info,
> >>>>>> and 2) strips the binaries. So the IPA module *.so file installed
> >>>>>> isn't the one which the signature was calculated against. Then
> >>>>>> the signature check fails, and libcamera tries to run the IPA
> >>>>>> module isolated (in a sandbox), which doesn't work if the IPA
> >>>>>> module wasn't designed to run isolated. The easiest way to fix that
> >>>>>> is to disable splitting out debug information and stripping the binaries
> >>>>>> during packaging with INHIBIT_PACKAGE_DEBUG_SPLIT and
> >>>>>> INHIBIT_PACKAGE_STRIP.
> >>>>>
> >>>>> This sounds like an effective solution for openembedded, but it needs to
> >>>>> be fixed in libcamera all the same.
> >>>>>
> >>>>> I'll try to follow up with the meson guys to see what we can do,.
> >>>
> >>> We re-sign the IPA modules at install time for this very specific
> >>> reason. If openembedded modifies the binaries after installing them,
> >>> should it re-run the signing script ?
> >>>
> >>>>>> Signed-off-by: Andrey Konovalov <andrey.konovalov at linaro.org>
> >>>>>> ---
> >>>>>>     .../recipes-multimedia/libcamera/libcamera.bb            | 9 ++++++++-
> >>>>>>     1 file changed, 8 insertions(+), 1 deletion(-)
> >>>>>>
> >>>>>> diff --git a/meta-multimedia/recipes-multimedia/libcamera/libcamera.bb b/meta-multimedia/recipes-multimedia/libcamera/libcamera.bb
> >>>>>> index 00a5c480d..573366f08 100644
> >>>>>> --- a/meta-multimedia/recipes-multimedia/libcamera/libcamera.bb
> >>>>>> +++ b/meta-multimedia/recipes-multimedia/libcamera/libcamera.bb
> >>>>>> @@ -18,13 +18,20 @@ PV = "202006+git${SRCPV}"
> >>>>>>     
> >>>>>>     S = "${WORKDIR}/git"
> >>>>>>     
> >>>>>> -DEPENDS = "python3-pyyaml-native udev gnutls boost"
> >>>>>> +DEPENDS = "python3-pyyaml-native udev gnutls boost chrpath-native"
> >>>>>>     DEPENDS += "${@bb.utils.contains('DISTRO_FEATURES', 'qt', 'qtbase qtbase-native', '', d)}"
> >>>>>>     
> >>>>>>     RDEPENDS_${PN} = "${@bb.utils.contains('DISTRO_FEATURES', 'wayland qt', 'qtwayland', '', d)}"
> >>>>>>     
> >>>>>>     inherit meson pkgconfig python3native
> >>>>>>     
> >>>>>> +do_install_append() {
> >>>>>> +        chrpath -d ${D}${libdir}/libcamera.so
> >>>>>
> >>>>> Aha, I didn't know about chrpath, that looks helpful. Perhaps part of
> >>>>> the solution will be handling our own strip/install actions to do this
> >>>>> explicitly in the build.
> >>>>>
> >>>>> It will be a pain to have to pull in another external dependency though...
> >>>>>
> >>>>>> +}
> >>>>>> +
> >>>>>>     FILES_${PN}-dev = "${includedir} ${libdir}/pkgconfig"
> >>>>>>     FILES_${PN} += " ${libdir}/libcamera.so"
> >>>>>>     
> >>>>>> +INHIBIT_PACKAGE_DEBUG_SPLIT = "1"
> >>>>>> +INHIBIT_PACKAGE_STRIP = "1"
> >>>>>> +

-- 
Regards,

Laurent Pinchart