lib/oe/package: Add strip keep-section support
v4 cover lib/oe/package: Add strip keep-section support
Hello,
This is v4 of the .debug_frame / keep-section patch, picking back up the thread from last year:
v1 (PACKAGE_KEEP_DEBUG_FRAME):
https://lists.openembedded.org/g/openembedded-core/message/203797
v2 (PACKAGE_KEEP_SECTIONS):
https://patchwork.yoctoproject.org/project/oe-core/patch/[email protected]/
Quick recap of the motivation: on 32-bit Arm, GCC's default EHABI unwind tables (.ARM.exidx / .ARM.extab) don't currently allow libunwind to produce a full backtrace once execution reaches the C++ termination path
(https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117941). GCC also emits DWARF Call Frame Information for the same functions under .debug_frame, which does allow a complete backtrace, but that section is unconditionally stripped today. This series lets a recipe opt back in on a per-section basis via PACKAGE_KEEP_SECTIONS, e.g.:
PACKAGE_KEEP_SECTIONS:pn-myrecipe = ".debug_frame"
What's new in v4:
- Fixed the new oe-selftest case (package.PackageKeepSections) itself:
it located busybox.nosuid with a hardcoded "usr/bin" prefix, which only exists when the usrmerge distro feature is enabled.
https://autobuilder.yoctoproject.org/valkyrie/#/builders/48/builds/4649 https://autobuilder.yoctoproject.org/valkyrie/#/builders/35/builds/4842
What's new in v3:
- Rebased on master, in particular on top of the runstrip() argument cleanup that landed since v2 -- no more tuple unpacking
- Added an oe-selftest case (package.PackageKeepSections) that builds core-image-minimal for qemuarm and uses readelf to show busybox loses .debug_frame by default and keeps it once PACKAGE_KEEP_SECTIONS is set, per Alexander's request for test coverage on v2.
- A couple of open points from the v2 thread I'd like to address before this goes further:
* Richard asked whether this should just be folded into minidebuginfo instead of being a separate "magic" option. minidebuginfo only injects compressed *symbol* information (.gnu_debugdata); it says nothing about unwind data. PACKAGE_KEEP_SECTIONS is orthogonal -- you need both symbols and unwind information to get a readable on-target backtrace, but a user may reasonably want only one of the two (e.g. GDB against a remote symbol store vs. a fully standalone coredump). I'd rather keep this as a separate, generic, opt-in knob than silently grow every minidebuginfo image by ~70KB per Arm binary that uses C++ exceptions.
* Khem asked whether minidebuginfo could reuse .ARM.exidx/.ARM.extab the way it uses .eh_frame elsewhere. Per the GCC bug above, those EHABI tables are exactly the ones that currently produce partial backtraces on Arm -- that limitation is the reason this series exists. If that gets fixed on the GCC side, PACKAGE_KEEP_SECTIONS simply becomes unnecessary for this particular use case, but until then .debug_frame is the only way to get complete backtraces on 32-bit Arm.
- The matching ref-manual documentation for PACKAGE_KEEP_SECTIONS has been sent as a separate patch to the docs list.
Mathieu Othacehe (1):
lib/oe/package: Add strip keep-section support
meta/classes-global/staging.bbclass | 4 +- meta/lib/oe/package.py | 19 +++++++-- meta/lib/oeqa/selftest/cases/package.py | 57 ++++++++++++++++++++++++- 3 files changed, 74 insertions(+), 6 deletions(-)
--
2.34.1
PATCH v4 1/1 lib/oe/package: Add strip keep-section support
On 32-bit Arm, the .ARM.extab and .ARM.exidx unwinding sections do not always provide enough information to get a full backtrace on C++ exceptions:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117941
In addition to those unwinding sections, GCC also emits unwinding instructions in DWARF format under the .debug_frame section. By instructing 'strip' not to remove that section, libunwind can use it to produce a full backtrace, for example when an unhandled C++ exception is thrown.
Add a PACKAGE_KEEP_SECTIONS variable, listing space separated ELF section names that should be kept around instead of stripped, for example:
PACKAGE_KEEP_SECTIONS:pn-myrecipe = ".debug_frame"
This variable is never applied to kernel modules, which are covered by CONFIG_UNWINDER_* instead.
Add a oe-selftest case building core-image-minimal for qemuarm, and checking with readelf that busybox loses its .debug_frame section by default, but keeps it once PACKAGE_KEEP_SECTIONS is set.
--- meta/classes-global/staging.bbclass | 4 +- meta/lib/oe/package.py | 19 +++++++-- meta/lib/oeqa/selftest/cases/package.py | 53 ++++++++++++++++++++++++- 3 files changed, 70 insertions(+), 6 deletions(-)
meta/classes-global/staging.bbclass +3 -1
@@ -91,10 +91,12 @@ python sysroot_strip () { base_libdir = d.getVar("base_libdir") qa_already_stripped = 'already-stripped' in (d.getVar('INSANE_SKIP:' + pn) or "").split() strip_cmd = d.getVar("STRIP") + keep_sections = d.getVar('PACKAGE_KEEP_SECTIONS') or "" max_process = oe.utils.get_bb_number_threads(d) oe.package.strip_execs(pn, dstdir, strip_cmd, libdir, base_libdir, max_process, - qa_already_stripped=qa_already_stripped) + qa_already_stripped=qa_already_stripped, + keep_sections=keep_sections) } do_populate_sysroot[dirs] = "${SYSROOT_DESTDIR}"
meta/lib/oe/package.py +15 -4
@@ -19,7 +19,7 @@ import shutil import bb.parse import oe.cachedpath -def runstrip(file, elftype, strip, extra_strip_sections=''): +def runstrip(file, elftype, strip, extra_strip_sections='', keep_sections=''): # Function to strip a single file, called from split_and_strip_files below # A working 'file' (one which works on the target architecture) #
@@ -49,6 +49,10 @@ def runstrip(file, elftype, strip, extra_strip_sections=''): for section in extra_strip_sections.split(): stripcmd.extend(["--remove-section=" + section]) + if keep_sections != '' and not elftype & 16: + for section in keep_sections.split(): + stripcmd.extend(["--keep-section=" + section]) + stripcmd.append(file) bb.debug(1, "runstrip: %s" % stripcmd)
@@ -96,7 +100,8 @@ def is_static_lib(path): return start == magic return False -def strip_execs(pn, dstdir, strip_cmd, libdir, base_libdir, max_process, qa_already_stripped=False): +def strip_execs(pn, dstdir, strip_cmd, libdir, base_libdir, max_process, + qa_already_stripped=False, keep_sections=''): """ Strip executable code (like executables, shared libraries) _in_place_ - Based on sysroot_strip in staging.bbclass
@@ -107,6 +112,9 @@ def strip_execs(pn, dstdir, strip_cmd, libdir, base_libdir, max_process, qa_alre :param max_process: number of stripping processes started in parallel :param qa_already_stripped: Set to True if already-stripped' in ${INSANE_SKIP} This is for proper logging and messages only. + :param keep_sections: Space separated list of ELF sections to keep even + though the file is being stripped, for example ".debug_frame" so that + libunwind can use it to generate backtraces. """ import stat, errno, oe.path, oe.utils
@@ -175,7 +183,8 @@ def strip_execs(pn, dstdir, strip_cmd, libdir, base_libdir, max_process, qa_alre elf_file = int(elffiles[file]) sfiles.append((file, elf_file, strip_cmd)) - oe.utils.multiprocess_launch_mp(runstrip, sfiles, max_process) + oe.utils.multiprocess_launch_mp(runstrip, sfiles, max_process, + extraargs=('', keep_sections)) TRANSLATE = ( ("@", "@at@"),
@@ -1359,7 +1368,9 @@ def process_split_and_strip_files(d): for f in staticlibs: sfiles.append((f, 16, strip)) - oe.utils.multiprocess_launch(oe.package.runstrip, sfiles, d) + keep_sections = d.getVar('PACKAGE_KEEP_SECTIONS') or "" + oe.utils.multiprocess_launch(oe.package.runstrip, sfiles, d, + extraargs=('', keep_sections)) # Build "minidebuginfo" and reinject it back into the stripped binaries if bb.utils.contains('DISTRO_FEATURES', 'minidebuginfo', True, False, d):
meta/lib/oeqa/selftest/cases/package.py +52 -1
65 lines not on this page · raw message
Signature · 1 line
2.34.1